Migrate workshop to agent-framework==1.2.1 (incl. native HandoffBuilder) by Copilot · Pull Request #419 · microsoft/OpenAIWorkshop

Copilot · 2026-04-28T16:39:41Z

Migrates the workshop from agent-framework==1.0.0rc1 to agent-framework==1.2.1, including a deeper rewrite of the multi-domain handoff agent on top of the now-native HandoffBuilder orchestration.

Investigation: what changes between 1.0.0rc1 and 1.2.1

Verified pip install agent-framework==1.2.1 succeeds and inspected the installed package source. Three things matter for this codebase:

Area	1.0.0rc1	1.2.1
Chat client	`agent_framework.azure.AzureOpenAIChatClient(deployment_name=…, endpoint=…)`	`agent_framework.openai.OpenAIChatClient(model=…, azure_endpoint=…)` — single unified client; auto-routes to Azure when `azure_endpoint` (or `credential`) is passed
`CheckpointStorage` protocol	`save_checkpoint`, `load_checkpoint`, `delete_checkpoint`; `WorkflowCheckpoint.workflow_id`	`save`, `load`, `delete`, `get_latest`; listing methods now require keyword-only `workflow_name`; `WorkflowCheckpoint.workflow_name`
Handoff orchestration	not in framework — workshop hand-rolled an intent classifier + regex pattern matching + manual context-prefix builder (~782-line agent)	first-class `agent_framework_orchestrations.HandoffBuilder` — auto-injects `handoff_to_<target>` tools per agent, intercepts via middleware, broadcasts shared conversation, emits `handoff_sent` events, supports checkpointing

Other APIs in active use (Agent, AgentSession, ChatOptions, MCPStreamableHTTPTool, MagenticBuilder, WorkflowEvent, etc.) remain compatible.

What this PR does

Dependency pins (4 files)

agentic_ai/applications/pyproject.toml, agentic_ai/applications/requirements.txt (regenerated with uv pip compile)
agentic_ai/workflow/fraud_detection_durable/pyproject.toml
agentic_ai/agents/mcp_agent_demo/pyproject.toml — replaces the older agent-framework-core>=0.2.0 + agent-framework-orchestrations>=0.1.0 split with the unified agent-framework==1.2.1 umbrella

Mechanical chat-client migration (13 files)
from agent_framework.azure import AzureOpenAIChatClient → from agent_framework.openai import OpenAIChatClient, with deployment_name= → model= and endpoint= → azure_endpoint=. Affects single_agent.py, magentic_group.py, reflection_agent.py, all 7 mcp_agent_demo modules, fraud_detection_durable/{worker,fraud_analysis_workflow}.py, and observability/sample_agent_with_tracing.py.

Deeper migration: handoff agent rewrite
agentic_ai/agents/agent_framework/multi_agent/handoff_multi_domain_agent.py is rewritten on top of HandoffBuilder. The hand-rolled intent classifier (Pydantic structured output), regex-based handoff phrase detection, and manual _build_context_prefix are gone. The new flow:

One Agent per domain, configured with a filtered MCP tool set and require_per_service_call_history_persistence=True (required by HandoffBuilder).
HandoffBuilder with default mesh topology — every specialist can hand off to every other specialist via auto-generated handoff_to_<target> tools that the LLM decides to call.
Cross-request continuity via with_checkpointing(_DictCheckpointStorage), with the storage backed by the per-session state_store. First turn → workflow.run(prompt, stream=True); subsequent turns → workflow.run(responses={pending_id: HandoffAgentUserRequest.create_response(prompt)}, checkpoint_id=…, stream=True).
WebSocket streaming preserved: output events with AgentResponseUpdate data drive agent_token/tool_called; handoff_sent events drive handoff_announcement + a follow-up agent_start. Synthetic handoff_to_* tool calls are filtered from the UI.

The new file is ~510 lines vs ~782 lines before, removing ~200 lines of regex / classifier code.

Checkpoint storage
Both _DictCheckpointStorage (handoff) and DictCheckpointStorage (magentic_group) updated to the 1.2.x CheckpointStorage protocol. magentic_group._purge_checkpoint_storage updated to use delete and the new keyword-only workflow_name.

Tests
tests/test_agent_framework_rc1_regression.py → tests/test_agent_framework_1_2_1_regression.py. Added coverage for the unified OpenAIChatClient (Azure routing, kwarg names), the native HandoffBuilder integration, the new CheckpointStorage protocol shape, and a save/load/get_latest roundtrip on DictCheckpointStorage. All 61 tests pass.

Docs

agentic_ai/agents/mcp_agent_demo/README.md — dependency table updated to the unified package.
agentic_ai/agents/agent_framework/multi_agent/HANDOFF_README.md — rewritten end-to-end to describe the new tool-based handoff flow (architecture diagram, configuration, sample multi-turn trace).

Validation

pytest tests/test_agent_framework_1_2_1_regression.py → 61 passed against agent-framework==1.2.1.
All migrated modules import cleanly with no fallback to legacy attributes.
GPT-5.5 review (run via the task subagent with model=gpt-5.5) surfaced 4 substantive bugs that an LLM-driven review would have shipped if not caught:
1. _DictCheckpointStorage implemented obsolete save_checkpoint/load_checkpoint and referenced removed WorkflowCheckpoint.workflow_id.
2. DictCheckpointStorage (magentic_group) had the same protocol drift, plus a stale delete_checkpoint reference in _purge_checkpoint_storage.
3. reflection_agent.py built OpenAIChatClient(**client_kwargs) from a dict whose keys still said "deployment_name"/"endpoint" (the migration regex matched only Python kwarg syntax).
4. observability/sample_agent_with_tracing.py still imported the removed ChatAgent, used the old chat_client=/thread/run_stream API, and had two undefined locals from a regex-misfire.
All four issues are fixed in this PR, and new regression tests lock in the protocol shape so they cannot silently regress.
parallel_validation (Code Review + CodeQL): CodeQL clean (0 alerts); 7 minor PEP 8 spacing comments in mcp_agent_demo modules addressed.

Things to be aware of

Native handoff is agent-driven: each specialist's instructions explicitly tell it to call the appropriate handoff_to_<target> tool when a request is out of domain. The legacy HANDOFF_LAZY_CLASSIFICATION and HANDOFF_CONTEXT_TRANSFER_TURNS env vars are no longer used (workflow shares conversation between specialists automatically).
The per-session checkpoint state lives at state_store[f"{session_id}_handoff_state"] and is trimmed to the last 5 checkpoints.
No live integration test against Azure OpenAI was run (no credentials in this environment); the regression suite covers the API surface, the agents' construction, the workflow event shape, and the checkpoint storage roundtrip, but real end-to-end MCP + Azure OpenAI behaviour should be smoke-tested manually before merge.

* update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * Updating Durable Agent Implementation (#404) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * chore: reduce anomaly probability to 1% for controlled demo pace * Reduce anomaly probability to 1% for controlled demo pace (#406) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * chore: reduce anomaly probability to 1% for controlled demo pace --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * add mcp_agent_demo * Add hybrid MCP server (strict-schema + natural-language tools) and typed-contract workflow Scripts 6-8 for the MCP agent demo: - workflow_typed_contracts.py: 4-agent IT security incident response pipeline with Pydantic-enforced contracts at every boundary (SecurityAlert, ThreatAssessment, ImpactAnalysis, IncidentResponse) - mcp_server_hybrid.py: Single MCP endpoint exposing BOTH tool types: * Strict-schema: triage_alert, assess_threat, create_response (Pydantic) * Natural-language: ask_security_advisor, explain_for_customer (prose) * Shared session state across all tools - mcp_client_hybrid.py: 5-step incident flow using both tool types with explicit context passing between steps - README.md: Updated from 6 to 8 capabilities with Hybrid Mode architecture diagram and Quick Start sections 7-8 * Replace workflow_local_remote.py with simplified proxy agent (Script 3) - Delete workflow_local_remote.py (old Script 3 with extra LLM call) - Add workflow_proxy_agent.py: MCPProxyAgent (BaseAgent, no LLM) calls MCP tools directly via call_tool() zero local LLM overhead - Simplify MCPProxyAgent: extract common _call() method, module-level _last_text() helper, remove verbose static methods - Fix workflow_typed_contracts.py: add async with context managers - Update README: merge Script 9 into Script 3 position, update architecture diagram, Quick Start, How It Works sections * Clean up MCP demos: remove comparison commentary, drop typed-contracts - workflow_proxy_agent.py: remove 'traditional vs proxy' framing MCPProxyAgent is the standard approach for remote agent integration - Delete workflow_typed_contracts.py: not relevant to MCP integration - README: 7 scripts, renumber hybrid 6-7, remove typed-contract sections * Add LangGraph + MAF GroupChat cross-framework demo (Scripts 8-9) - mcp_server_langgraph.py: LangGraph ReAct agent exposed as MCP server on port 8003 with architecture tools (pattern eval, migration estimate, tech stack recommendation) - workflow_group_chat.py: MAF GroupChatBuilder orchestration with LLM planner routing between local BusinessStrategist (MAF) and remote TechnicalArchitect (LangGraph via MCP) - pyproject.toml: add langgraph, langchain-openai, langchain-core deps - README.md: update to 9 scripts, add cross-framework architecture diagram, Quick Start sections, and dependency table entries * MCP agent demo: rewrite README as MCP-vs-A2A thesis, delete proxy agent, finalize group chat with inline Planner * README: Mermaid diagrams, professional tone, remove LinkedIn teasing language * README: reframe around multi-framework interop problem, two design patterns (Agent-as-Tool + Agent Adapter), add conceptual architecture * README: fix Mermaid diagrams - use br tags instead of \n for line breaks * README: fix A2A comparison table - elicitation supported, structured schemas not * Upgrade agent-framework to 1.0.0rc1 and fastmcp to 3.0.2 Breaking changes migrated: - agent-framework 1.0.0rc1: ChatAgent->Agent, AgentThread->AgentSession, run_stream->run(stream=True), model->default_options(ChatOptions), WorkflowOutputEvent/AgentRunEvent/RequestInfoEvent->unified WorkflowEvent, MagenticBuilder now uses constructor kwargs, orchestrations moved to agent_framework_orchestrations package - fastmcp 3.0.2: removed mcp version pin (now pulled by fastmcp), no code changes needed (fully compatible) Files updated: - agentic_ai/agents/agent_framework/single_agent.py - agentic_ai/agents/agent_framework/multi_agent/handoff_multi_domain_agent.py - agentic_ai/agents/agent_framework/multi_agent/reflection_agent.py - agentic_ai/agents/agent_framework/multi_agent/magentic_group.py - agentic_ai/applications/pyproject.toml (agent-framework==1.0.0rc1) - agentic_ai/workflow/fraud_detection_durable/pyproject.toml (agent-framework==1.0.0rc1) - mcp/pyproject.toml (fastmcp==3.0.2) - tests/test_agent_framework_rc1_regression.py (51 regression tests, all passing) * Fix CI: regenerate requirements.txt with agent-framework-core==1.0.0rc1 The Docker build uses requirements.txt (not pyproject.toml/uv.lock), and it still had agent-framework-core==1.0.0b260130 pinned, which conflicts with agent-framework==1.0.0rc1 requiring agent-framework-core==1.0.0rc1. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22315210980 * Fix: use GH_PAT for auto-merge to trigger downstream workflows Pushes made by GITHUB_TOKEN don't trigger other workflows (GitHub Actions limitation to prevent infinite loops). This meant the auto-merge into int-agentic never triggered promote-to-main.yml. Now uses secrets.GH_PAT with fallback to GITHUB_TOKEN so: 1. If GH_PAT is configured: merge triggers promote-to-main.yml 2. If not: merge still works, but promotion PR must be created manually * Add workflow_dispatch trigger to promote-to-main Allows manual triggering when auto-trigger is missed (e.g. after GITHUB_TOKEN-based merges that don't fire downstream workflows). * Fix: Cosmos DB auth + remove phantom reflection_workflow_agent 1. MCP Cosmos DB: Replace AzureCliCredential (not available in containers) with ManagedIdentityCredential (when AZURE_CLIENT_ID is set) or DefaultAzureCredential as fallback. This fixes the 'Azure CLI not found on path' error in production Container Apps. 2. Backend: Remove reflection_workflow_agent from DEFAULT_AVAILABLE_AGENTS and descriptions - the module file doesn't exist, causing it to show as a broken option in the UI agent selector. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22321497108 * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Trigger rebuild: deploy Cosmos DB auth fix and agent list cleanup * Fix: iterate ResponseStream directly, not .updates ResponseStream is AsyncIterable - iterate with 'async for chunk in stream:' The .updates property returns a Sequence (list), not an async iterator. This caused TypeError: 'async for' requires __aiter__, got list. Not caught in local tests because unit tests mocked the streaming and the regression tests verified API signatures but didn't run live agents. * Fix: make integration test failures block the pipeline Removed continue-on-error: true from the pytest step so test failures actually fail the workflow. Previously, 5/7 tests could fail and the pipeline would still report success and proceed to auto-merge. * Fix: use SHA-tagged images for container deployment, not :latest Container Apps don't create a new revision when the image tag is unchanged. Using :latest meant the old container kept running even after a new image was pushed. Now passes github.sha as the image tag, which forces a new revision on every deploy. This was the root cause of the ResponseStream.updates bug reaching production - the integration tests ran against old containers. * Skip MCP localhost check in CI mode In CI, the evaluation talks to the backend via HTTP which connects to the deployed MCP container internally. The localhost:8000 check is only useful for local dev and was producing a confusing warning in CI logs. * Fix: auto-merge creates PR if none exists Previously auto-merge only looked for an existing open PR and skipped if none was found. Now it auto-creates a PR from the dev branch to int-agentic if one doesn't exist, then merges it. This ensures every successful pipeline run promotes code to int-agentic regardless of whether a PR was manually created beforehand. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22330447642 --------- Co-authored-by: James N. <james.nguyen@microsoft.com>

* update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * Updating Durable Agent Implementation (#404) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * chore: reduce anomaly probability to 1% for controlled demo pace * Reduce anomaly probability to 1% for controlled demo pace (#406) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * chore: reduce anomaly probability to 1% for controlled demo pace --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * add mcp_agent_demo * Add hybrid MCP server (strict-schema + natural-language tools) and typed-contract workflow Scripts 6-8 for the MCP agent demo: - workflow_typed_contracts.py: 4-agent IT security incident response pipeline with Pydantic-enforced contracts at every boundary (SecurityAlert, ThreatAssessment, ImpactAnalysis, IncidentResponse) - mcp_server_hybrid.py: Single MCP endpoint exposing BOTH tool types: * Strict-schema: triage_alert, assess_threat, create_response (Pydantic) * Natural-language: ask_security_advisor, explain_for_customer (prose) * Shared session state across all tools - mcp_client_hybrid.py: 5-step incident flow using both tool types with explicit context passing between steps - README.md: Updated from 6 to 8 capabilities with Hybrid Mode architecture diagram and Quick Start sections 7-8 * Replace workflow_local_remote.py with simplified proxy agent (Script 3) - Delete workflow_local_remote.py (old Script 3 with extra LLM call) - Add workflow_proxy_agent.py: MCPProxyAgent (BaseAgent, no LLM) calls MCP tools directly via call_tool() zero local LLM overhead - Simplify MCPProxyAgent: extract common _call() method, module-level _last_text() helper, remove verbose static methods - Fix workflow_typed_contracts.py: add async with context managers - Update README: merge Script 9 into Script 3 position, update architecture diagram, Quick Start, How It Works sections * Clean up MCP demos: remove comparison commentary, drop typed-contracts - workflow_proxy_agent.py: remove 'traditional vs proxy' framing MCPProxyAgent is the standard approach for remote agent integration - Delete workflow_typed_contracts.py: not relevant to MCP integration - README: 7 scripts, renumber hybrid 6-7, remove typed-contract sections * Add LangGraph + MAF GroupChat cross-framework demo (Scripts 8-9) - mcp_server_langgraph.py: LangGraph ReAct agent exposed as MCP server on port 8003 with architecture tools (pattern eval, migration estimate, tech stack recommendation) - workflow_group_chat.py: MAF GroupChatBuilder orchestration with LLM planner routing between local BusinessStrategist (MAF) and remote TechnicalArchitect (LangGraph via MCP) - pyproject.toml: add langgraph, langchain-openai, langchain-core deps - README.md: update to 9 scripts, add cross-framework architecture diagram, Quick Start sections, and dependency table entries * MCP agent demo: rewrite README as MCP-vs-A2A thesis, delete proxy agent, finalize group chat with inline Planner * README: Mermaid diagrams, professional tone, remove LinkedIn teasing language * README: reframe around multi-framework interop problem, two design patterns (Agent-as-Tool + Agent Adapter), add conceptual architecture * README: fix Mermaid diagrams - use br tags instead of \n for line breaks * README: fix A2A comparison table - elicitation supported, structured schemas not * Upgrade agent-framework to 1.0.0rc1 and fastmcp to 3.0.2 Breaking changes migrated: - agent-framework 1.0.0rc1: ChatAgent->Agent, AgentThread->AgentSession, run_stream->run(stream=True), model->default_options(ChatOptions), WorkflowOutputEvent/AgentRunEvent/RequestInfoEvent->unified WorkflowEvent, MagenticBuilder now uses constructor kwargs, orchestrations moved to agent_framework_orchestrations package - fastmcp 3.0.2: removed mcp version pin (now pulled by fastmcp), no code changes needed (fully compatible) Files updated: - agentic_ai/agents/agent_framework/single_agent.py - agentic_ai/agents/agent_framework/multi_agent/handoff_multi_domain_agent.py - agentic_ai/agents/agent_framework/multi_agent/reflection_agent.py - agentic_ai/agents/agent_framework/multi_agent/magentic_group.py - agentic_ai/applications/pyproject.toml (agent-framework==1.0.0rc1) - agentic_ai/workflow/fraud_detection_durable/pyproject.toml (agent-framework==1.0.0rc1) - mcp/pyproject.toml (fastmcp==3.0.2) - tests/test_agent_framework_rc1_regression.py (51 regression tests, all passing) * Fix CI: regenerate requirements.txt with agent-framework-core==1.0.0rc1 The Docker build uses requirements.txt (not pyproject.toml/uv.lock), and it still had agent-framework-core==1.0.0b260130 pinned, which conflicts with agent-framework==1.0.0rc1 requiring agent-framework-core==1.0.0rc1. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22315210980 * Fix: use GH_PAT for auto-merge to trigger downstream workflows Pushes made by GITHUB_TOKEN don't trigger other workflows (GitHub Actions limitation to prevent infinite loops). This meant the auto-merge into int-agentic never triggered promote-to-main.yml. Now uses secrets.GH_PAT with fallback to GITHUB_TOKEN so: 1. If GH_PAT is configured: merge triggers promote-to-main.yml 2. If not: merge still works, but promotion PR must be created manually * Add workflow_dispatch trigger to promote-to-main Allows manual triggering when auto-trigger is missed (e.g. after GITHUB_TOKEN-based merges that don't fire downstream workflows). * Fix: Cosmos DB auth + remove phantom reflection_workflow_agent 1. MCP Cosmos DB: Replace AzureCliCredential (not available in containers) with ManagedIdentityCredential (when AZURE_CLIENT_ID is set) or DefaultAzureCredential as fallback. This fixes the 'Azure CLI not found on path' error in production Container Apps. 2. Backend: Remove reflection_workflow_agent from DEFAULT_AVAILABLE_AGENTS and descriptions - the module file doesn't exist, causing it to show as a broken option in the UI agent selector. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22321497108 * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Trigger rebuild: deploy Cosmos DB auth fix and agent list cleanup * Fix: iterate ResponseStream directly, not .updates ResponseStream is AsyncIterable - iterate with 'async for chunk in stream:' The .updates property returns a Sequence (list), not an async iterator. This caused TypeError: 'async for' requires __aiter__, got list. Not caught in local tests because unit tests mocked the streaming and the regression tests verified API signatures but didn't run live agents. * Fix: make integration test failures block the pipeline Removed continue-on-error: true from the pytest step so test failures actually fail the workflow. Previously, 5/7 tests could fail and the pipeline would still report success and proceed to auto-merge. * Fix: use SHA-tagged images for container deployment, not :latest Container Apps don't create a new revision when the image tag is unchanged. Using :latest meant the old container kept running even after a new image was pushed. Now passes github.sha as the image tag, which forces a new revision on every deploy. This was the root cause of the ResponseStream.updates bug reaching production - the integration tests ran against old containers. * Skip MCP localhost check in CI mode In CI, the evaluation talks to the backend via HTTP which connects to the deployed MCP container internally. The localhost:8000 check is only useful for local dev and was producing a confusing warning in CI logs. * Fix: auto-merge creates PR if none exists Previously auto-merge only looked for an existing open PR and skipped if none was found. Now it auto-creates a PR from the dev branch to int-agentic if one doesn't exist, then merges it. This ensures every successful pipeline run promotes code to int-agentic regardless of whether a PR was manually created beforehand. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22330447642 * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22334055353 --------- Co-authored-by: James N. <james.nguyen@microsoft.com>

…nAIChatClient Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com>

…ts for 1.2.1 Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com>

…migration Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com>

…ent kwargs, observability sample Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com>

Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com>

james-tn · 2026-04-28T17:08:53Z

@copilot I re-checked this after the base branch changed to int-agentic. I still see two issues that look worth fixing before merge:

High - uv.lock files still pin the old Agent Framework packages. The migrated pyproject.toml files now require agent-framework==1.2.1, but the committed lockfiles were not regenerated, so the documented uv sync / uv run paths can still install the old APIs. Examples:
- agentic_ai/applications/pyproject.toml:8 says agent-framework==1.2.1, but agentic_ai/applications/uv.lock:42 still locks agent-framework to 1.0.0rc1, and agentic_ai/applications/uv.lock:586 still records the project specifier as ==1.0.0rc1.
- agentic_ai/workflow/fraud_detection_durable/pyproject.toml:9 says agent-framework==1.2.1, but agentic_ai/workflow/fraud_detection_durable/uv.lock:42 still locks 1.0.0rc1, and agentic_ai/workflow/fraud_detection_durable/uv.lock:1098 still has the older durabletask specifier.
- agentic_ai/agents/mcp_agent_demo/pyproject.toml:8 switches to the umbrella agent-framework==1.2.1, but agentic_ai/agents/mcp_agent_demo/uv.lock:896-897 still records the old split agent-framework-core / agent-framework-orchestrations specs.
Medium - custom Magentic checkpoint storage resume can silently fail under the 1.2.x protocol. DictCheckpointStorage.list_checkpoints, list_checkpoint_ids, and get_latest now require keyword-only workflow_name, but _get_latest_checkpoint_id() still calls the fallback methods without that keyword at agentic_ai/agents/agent_framework/multi_agent/magentic_group.py:1127 and agentic_ai/agents/agent_framework/multi_agent/magentic_group.py:1141. Those TypeErrors are swallowed, so custom storage implementations that follow only the 1.2.x protocol and do not expose the nonstandard latest_checkpoint_id property will fail to find resumable checkpoints.

Validation notes: confirmed PR 419 now targets int-agentic; earlier git diff --check origin/main...origin/pr/419 was clean, and I re-inspected the affected dependency/checkpoint paths on the PR branch after the base retarget.

Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/91307ce7-2444-4b5f-b75f-2b736bc022c5 Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com>

Copilot · 2026-04-28T17:12:17Z

@copilot I re-checked this after the base branch changed to int-agentic. I still see two issues that look worth fixing before merge:

**High - uv.lock files still pin the old Agent Framework pack...

Both fixed in 038dfb6:

Regenerated all three uv.lock files with uv lock. agent-framework is now 1.2.1 in agentic_ai/applications/uv.lock, agentic_ai/workflow/fraud_detection_durable/uv.lock, and agentic_ai/agents/mcp_agent_demo/uv.lock (the last one also dropped the split agent-framework-core / agent-framework-orchestrations specifiers in favor of the umbrella package).
_get_latest_checkpoint_id now resolves the workflow name via _workflow_name_for_storage and passes workflow_name= to get_latest / list_checkpoints / list_checkpoint_ids. It also tries get_latest first (the 1.2.x shortcut) before falling back to listing. Added test_get_latest_checkpoint_id_uses_workflow_name_kwarg, which exercises a storage exposing only the 1.2.x protocol; full suite is 62/62 green.

* update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * Updating Durable Agent Implementation (#404) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * chore: reduce anomaly probability to 1% for controlled demo pace * Reduce anomaly probability to 1% for controlled demo pace (#406) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * chore: reduce anomaly probability to 1% for controlled demo pace --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * add mcp_agent_demo * Add hybrid MCP server (strict-schema + natural-language tools) and typed-contract workflow Scripts 6-8 for the MCP agent demo: - workflow_typed_contracts.py: 4-agent IT security incident response pipeline with Pydantic-enforced contracts at every boundary (SecurityAlert, ThreatAssessment, ImpactAnalysis, IncidentResponse) - mcp_server_hybrid.py: Single MCP endpoint exposing BOTH tool types: * Strict-schema: triage_alert, assess_threat, create_response (Pydantic) * Natural-language: ask_security_advisor, explain_for_customer (prose) * Shared session state across all tools - mcp_client_hybrid.py: 5-step incident flow using both tool types with explicit context passing between steps - README.md: Updated from 6 to 8 capabilities with Hybrid Mode architecture diagram and Quick Start sections 7-8 * Replace workflow_local_remote.py with simplified proxy agent (Script 3) - Delete workflow_local_remote.py (old Script 3 with extra LLM call) - Add workflow_proxy_agent.py: MCPProxyAgent (BaseAgent, no LLM) calls MCP tools directly via call_tool() zero local LLM overhead - Simplify MCPProxyAgent: extract common _call() method, module-level _last_text() helper, remove verbose static methods - Fix workflow_typed_contracts.py: add async with context managers - Update README: merge Script 9 into Script 3 position, update architecture diagram, Quick Start, How It Works sections * Clean up MCP demos: remove comparison commentary, drop typed-contracts - workflow_proxy_agent.py: remove 'traditional vs proxy' framing MCPProxyAgent is the standard approach for remote agent integration - Delete workflow_typed_contracts.py: not relevant to MCP integration - README: 7 scripts, renumber hybrid 6-7, remove typed-contract sections * Add LangGraph + MAF GroupChat cross-framework demo (Scripts 8-9) - mcp_server_langgraph.py: LangGraph ReAct agent exposed as MCP server on port 8003 with architecture tools (pattern eval, migration estimate, tech stack recommendation) - workflow_group_chat.py: MAF GroupChatBuilder orchestration with LLM planner routing between local BusinessStrategist (MAF) and remote TechnicalArchitect (LangGraph via MCP) - pyproject.toml: add langgraph, langchain-openai, langchain-core deps - README.md: update to 9 scripts, add cross-framework architecture diagram, Quick Start sections, and dependency table entries * MCP agent demo: rewrite README as MCP-vs-A2A thesis, delete proxy agent, finalize group chat with inline Planner * README: Mermaid diagrams, professional tone, remove LinkedIn teasing language * README: reframe around multi-framework interop problem, two design patterns (Agent-as-Tool + Agent Adapter), add conceptual architecture * README: fix Mermaid diagrams - use br tags instead of \n for line breaks * README: fix A2A comparison table - elicitation supported, structured schemas not * Upgrade agent-framework to 1.0.0rc1 and fastmcp to 3.0.2 Breaking changes migrated: - agent-framework 1.0.0rc1: ChatAgent->Agent, AgentThread->AgentSession, run_stream->run(stream=True), model->default_options(ChatOptions), WorkflowOutputEvent/AgentRunEvent/RequestInfoEvent->unified WorkflowEvent, MagenticBuilder now uses constructor kwargs, orchestrations moved to agent_framework_orchestrations package - fastmcp 3.0.2: removed mcp version pin (now pulled by fastmcp), no code changes needed (fully compatible) Files updated: - agentic_ai/agents/agent_framework/single_agent.py - agentic_ai/agents/agent_framework/multi_agent/handoff_multi_domain_agent.py - agentic_ai/agents/agent_framework/multi_agent/reflection_agent.py - agentic_ai/agents/agent_framework/multi_agent/magentic_group.py - agentic_ai/applications/pyproject.toml (agent-framework==1.0.0rc1) - agentic_ai/workflow/fraud_detection_durable/pyproject.toml (agent-framework==1.0.0rc1) - mcp/pyproject.toml (fastmcp==3.0.2) - tests/test_agent_framework_rc1_regression.py (51 regression tests, all passing) * Fix CI: regenerate requirements.txt with agent-framework-core==1.0.0rc1 The Docker build uses requirements.txt (not pyproject.toml/uv.lock), and it still had agent-framework-core==1.0.0b260130 pinned, which conflicts with agent-framework==1.0.0rc1 requiring agent-framework-core==1.0.0rc1. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22315210980 * Fix: use GH_PAT for auto-merge to trigger downstream workflows Pushes made by GITHUB_TOKEN don't trigger other workflows (GitHub Actions limitation to prevent infinite loops). This meant the auto-merge into int-agentic never triggered promote-to-main.yml. Now uses secrets.GH_PAT with fallback to GITHUB_TOKEN so: 1. If GH_PAT is configured: merge triggers promote-to-main.yml 2. If not: merge still works, but promotion PR must be created manually * Add workflow_dispatch trigger to promote-to-main Allows manual triggering when auto-trigger is missed (e.g. after GITHUB_TOKEN-based merges that don't fire downstream workflows). * Fix: Cosmos DB auth + remove phantom reflection_workflow_agent 1. MCP Cosmos DB: Replace AzureCliCredential (not available in containers) with ManagedIdentityCredential (when AZURE_CLIENT_ID is set) or DefaultAzureCredential as fallback. This fixes the 'Azure CLI not found on path' error in production Container Apps. 2. Backend: Remove reflection_workflow_agent from DEFAULT_AVAILABLE_AGENTS and descriptions - the module file doesn't exist, causing it to show as a broken option in the UI agent selector. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22321497108 * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Trigger rebuild: deploy Cosmos DB auth fix and agent list cleanup * Fix: iterate ResponseStream directly, not .updates ResponseStream is AsyncIterable - iterate with 'async for chunk in stream:' The .updates property returns a Sequence (list), not an async iterator. This caused TypeError: 'async for' requires __aiter__, got list. Not caught in local tests because unit tests mocked the streaming and the regression tests verified API signatures but didn't run live agents. * Fix: make integration test failures block the pipeline Removed continue-on-error: true from the pytest step so test failures actually fail the workflow. Previously, 5/7 tests could fail and the pipeline would still report success and proceed to auto-merge. * Fix: use SHA-tagged images for container deployment, not :latest Container Apps don't create a new revision when the image tag is unchanged. Using :latest meant the old container kept running even after a new image was pushed. Now passes github.sha as the image tag, which forces a new revision on every deploy. This was the root cause of the ResponseStream.updates bug reaching production - the integration tests ran against old containers. * Skip MCP localhost check in CI mode In CI, the evaluation talks to the backend via HTTP which connects to the deployed MCP container internally. The localhost:8000 check is only useful for local dev and was producing a confusing warning in CI logs. * Fix: auto-merge creates PR if none exists Previously auto-merge only looked for an existing open PR and skipped if none was found. Now it auto-creates a PR from the dev branch to int-agentic if one doesn't exist, then merges it. This ensures every successful pipeline run promotes code to int-agentic regardless of whether a PR was manually created beforehand. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22330447642 * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22334055353 * Migrate workshop to agent-framework==1.2.1 (incl. native HandoffBuilder) (#419) * Promote: int-agentic → main (production) (#414) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * Updating Durable Agent Implementation (#404) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * chore: reduce anomaly probability to 1% for controlled demo pace * Reduce anomaly probability to 1% for controlled demo pace (#406) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * chore: reduce anomaly probability to 1% for controlled demo pace --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * add mcp_agent_demo * Add hybrid MCP server (strict-schema + natural-language tools) and typed-contract workflow Scripts 6-8 for the MCP agent demo: - workflow_typed_contracts.py: 4-agent IT security incident response pipeline with Pydantic-enforced contracts at every boundary (SecurityAlert, ThreatAssessment, ImpactAnalysis, IncidentResponse) - mcp_server_hybrid.py: Single MCP endpoint exposing BOTH tool types: * Strict-schema: triage_alert, assess_threat, create_response (Pydantic) * Natural-language: ask_security_advisor, explain_for_customer (prose) * Shared session state across all tools - mcp_client_hybrid.py: 5-step incident flow using both tool types with explicit context passing between steps - README.md: Updated from 6 to 8 capabilities with Hybrid Mode architecture diagram and Quick Start sections 7-8 * Replace workflow_local_remote.py with simplified proxy agent (Script 3) - Delete workflow_local_remote.py (old Script 3 with extra LLM call) - Add workflow_proxy_agent.py: MCPProxyAgent (BaseAgent, no LLM) calls MCP tools directly via call_tool() zero local LLM overhead - Simplify MCPProxyAgent: extract common _call() method, module-level _last_text() helper, remove verbose static methods - Fix workflow_typed_contracts.py: add async with context managers - Update README: merge Script 9 into Script 3 position, update architecture diagram, Quick Start, How It Works sections * Clean up MCP demos: remove comparison commentary, drop typed-contracts - workflow_proxy_agent.py: remove 'traditional vs proxy' framing MCPProxyAgent is the standard approach for remote agent integration - Delete workflow_typed_contracts.py: not relevant to MCP integration - README: 7 scripts, renumber hybrid 6-7, remove typed-contract sections * Add LangGraph + MAF GroupChat cross-framework demo (Scripts 8-9) - mcp_server_langgraph.py: LangGraph ReAct agent exposed as MCP server on port 8003 with architecture tools (pattern eval, migration estimate, tech stack recommendation) - workflow_group_chat.py: MAF GroupChatBuilder orchestration with LLM planner routing between local BusinessStrategist (MAF) and remote TechnicalArchitect (LangGraph via MCP) - pyproject.toml: add langgraph, langchain-openai, langchain-core deps - README.md: update to 9 scripts, add cross-framework architecture diagram, Quick Start sections, and dependency table entries * MCP agent demo: rewrite README as MCP-vs-A2A thesis, delete proxy agent, finalize group chat with inline Planner * README: Mermaid diagrams, professional tone, remove LinkedIn teasing language * README: reframe around multi-framework interop problem, two design patterns (Agent-as-Tool + Agent Adapter), add conceptual architecture * README: fix Mermaid diagrams - use br tags instead of \n for line breaks * README: fix A2A comparison table - elicitation supported, structured schemas not * Upgrade agent-framework to 1.0.0rc1 and fastmcp to 3.0.2 Breaking changes migrated: - agent-framework 1.0.0rc1: ChatAgent->Agent, AgentThread->AgentSession, run_stream->run(stream=True), model->default_options(ChatOptions), WorkflowOutputEvent/AgentRunEvent/RequestInfoEvent->unified WorkflowEvent, MagenticBuilder now uses constructor kwargs, orchestrations moved to agent_framework_orchestrations package - fastmcp 3.0.2: removed mcp version pin (now pulled by fastmcp), no code changes needed (fully compatible) Files updated: - agentic_ai/agents/agent_framework/single_agent.py - agentic_ai/agents/agent_framework/multi_agent/handoff_multi_domain_agent.py - agentic_ai/agents/agent_framework/multi_agent/reflection_agent.py - agentic_ai/agents/agent_framework/multi_agent/magentic_group.py - agentic_ai/applications/pyproject.toml (agent-framework==1.0.0rc1) - agentic_ai/workflow/fraud_detection_durable/pyproject.toml (agent-framework==1.0.0rc1) - mcp/pyproject.toml (fastmcp==3.0.2) - tests/test_agent_framework_rc1_regression.py (51 regression tests, all passing) * Fix CI: regenerate requirements.txt with agent-framework-core==1.0.0rc1 The Docker build uses requirements.txt (not pyproject.toml/uv.lock), and it still had agent-framework-core==1.0.0b260130 pinned, which conflicts with agent-framework==1.0.0rc1 requiring agent-framework-core==1.0.0rc1. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22315210980 * Fix: use GH_PAT for auto-merge to trigger downstream workflows Pushes made by GITHUB_TOKEN don't trigger other workflows (GitHub Actions limitation to prevent infinite loops). This meant the auto-merge into int-agentic never triggered promote-to-main.yml. Now uses secrets.GH_PAT with fallback to GITHUB_TOKEN so: 1. If GH_PAT is configured: merge triggers promote-to-main.yml 2. If not: merge still works, but promotion PR must be created manually * Add workflow_dispatch trigger to promote-to-main Allows manual triggering when auto-trigger is missed (e.g. after GITHUB_TOKEN-based merges that don't fire downstream workflows). * Fix: Cosmos DB auth + remove phantom reflection_workflow_agent 1. MCP Cosmos DB: Replace AzureCliCredential (not available in containers) with ManagedIdentityCredential (when AZURE_CLIENT_ID is set) or DefaultAzureCredential as fallback. This fixes the 'Azure CLI not found on path' error in production Container Apps. 2. Backend: Remove reflection_workflow_agent from DEFAULT_AVAILABLE_AGENTS and descriptions - the module file doesn't exist, causing it to show as a broken option in the UI agent selector. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22321497108 * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Trigger rebuild: deploy Cosmos DB auth fix and agent list cleanup * Fix: iterate ResponseStream directly, not .updates ResponseStream is AsyncIterable - iterate with 'async for chunk in stream:' The .updates property returns a Sequence (list), not an async iterator. This caused TypeError: 'async for' requires __aiter__, got list. Not caught in local tests because unit tests mocked the streaming and the regression tests verified API signatures but didn't run live agents. * Fix: make integration test failures block the pipeline Removed continue-on-error: true from the pytest step so test failures actually fail the workflow. Previously, 5/7 tests could fail and the pipeline would still report success and proceed to auto-merge. * Fix: use SHA-tagged images for container deployment, not :latest Container Apps don't create a new revision when the image tag is unchanged. Using :latest meant the old container kept running even after a new image was pushed. Now passes github.sha as the image tag, which forces a new revision on every deploy. This was the root cause of the ResponseStream.updates bug reaching production - the integration tests ran against old containers. * Skip MCP localhost check in CI mode In CI, the evaluation talks to the backend via HTTP which connects to the deployed MCP container internally. The localhost:8000 check is only useful for local dev and was producing a confusing warning in CI logs. * Fix: auto-merge creates PR if none exists Previously auto-merge only looked for an existing open PR and skipped if none was found. Now it auto-creates a PR from the dev branch to int-agentic if one doesn't exist, then merges it. This ensures every successful pipeline run promotes code to int-agentic regardless of whether a PR was manually created beforehand. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22330447642 --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * Promote: int-agentic → main (production) (#416) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * Updating Durable Agent Implementation (#404) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * chore: reduce anomaly probability to 1% for controlled demo pace * Reduce anomaly probability to 1% for controlled demo pace (#406) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * chore: reduce anomaly probability to 1% for controlled demo pace --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * add mcp_agent_demo * Add hybrid MCP server (strict-schema + natural-language tools) and typed-contract workflow Scripts 6-8 for the MCP agent demo: - workflow_typed_contracts.py: 4-agent IT security incident response pipeline with Pydantic-enforced contracts at every boundary (SecurityAlert, ThreatAssessment, ImpactAnalysis, IncidentResponse) - mcp_server_hybrid.py: Single MCP endpoint exposing BOTH tool types: * Strict-schema: triage_alert, assess_threat, create_response (Pydantic) * Natural-language: ask_security_advisor, explain_for_customer (prose) * Shared session state across all tools - mcp_client_hybrid.py: 5-step incident flow using both tool types with explicit context passing between steps - README.md: Updated from 6 to 8 capabilities with Hybrid Mode architecture diagram and Quick Start sections 7-8 * Replace workflow_local_remote.py with simplified proxy agent (Script 3) - Delete workflow_local_remote.py (old Script 3 with extra LLM call) - Add workflow_proxy_agent.py: MCPProxyAgent (BaseAgent, no LLM) calls MCP tools directly via call_tool() zero local LLM overhead - Simplify MCPProxyAgent: extract common _call() method, module-level _last_text() helper, remove verbose static methods - Fix workflow_typed_contracts.py: add async with context managers - Update README: merge Script 9 into Script 3 position, update architecture diagram, Quick Start, How It Works sections * Clean up MCP demos: remove comparison commentary, drop typed-contracts - workflow_proxy_agent.py: remove 'traditional vs proxy' framing MCPProxyAgent is the standard approach for remote agent integration - Delete workflow_typed_contracts.py: not relevant to MCP integration - README: 7 scripts, renumber hybrid 6-7, remove typed-contract sections * Add LangGraph + MAF GroupChat cross-framework demo (Scripts 8-9) - mcp_server_langgraph.py: LangGraph ReAct agent exposed as MCP server on port 8003 with architecture tools (pattern eval, migration estimate, tech stack recommendation) - workflow_group_chat.py: MAF GroupChatBuilder orchestration with LLM planner routing between local BusinessStrategist (MAF) and remote TechnicalArchitect (LangGraph via MCP) - pyproject.toml: add langgraph, langchain-openai, langchain-core deps - README.md: update to 9 scripts, add cross-framework architecture diagram, Quick Start sections, and dependency table entries * MCP agent demo: rewrite README as MCP-vs-A2A thesis, delete proxy agent, finalize group chat with inline Planner * README: Mermaid diagrams, professional tone, remove LinkedIn teasing language * README: reframe around multi-framework interop problem, two design patterns (Agent-as-Tool + Agent Adapter), add conceptual architecture * README: fix Mermaid diagrams - use br tags instead of \n for line breaks * README: fix A2A comparison table - elicitation supported, structured schemas not * Upgrade agent-framework to 1.0.0rc1 and fastmcp to 3.0.2 Breaking changes migrated: - agent-framework 1.0.0rc1: ChatAgent->Agent, AgentThread->AgentSession, run_stream->run(stream=True), model->default_options(ChatOptions), WorkflowOutputEvent/AgentRunEvent/RequestInfoEvent->unified WorkflowEvent, MagenticBuilder now uses constructor kwargs, orchestrations moved to agent_framework_orchestrations package - fastmcp 3.0.2: removed mcp version pin (now pulled by fastmcp), no code changes needed (fully compatible) Files updated: - agentic_ai/agents/agent_framework/single_agent.py - agentic_ai/agents/agent_framework/multi_agent/handoff_multi_domain_agent.py - agentic_ai/agents/agent_framework/multi_agent/reflection_agent.py - agentic_ai/agents/agent_framework/multi_agent/magentic_group.py - agentic_ai/applications/pyproject.toml (agent-framework==1.0.0rc1) - agentic_ai/workflow/fraud_detection_durable/pyproject.toml (agent-framework==1.0.0rc1) - mcp/pyproject.toml (fastmcp==3.0.2) - tests/test_agent_framework_rc1_regression.py (51 regression tests, all passing) * Fix CI: regenerate requirements.txt with agent-framework-core==1.0.0rc1 The Docker build uses requirements.txt (not pyproject.toml/uv.lock), and it still had agent-framework-core==1.0.0b260130 pinned, which conflicts with agent-framework==1.0.0rc1 requiring agent-framework-core==1.0.0rc1. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22315210980 * Fix: use GH_PAT for auto-merge to trigger downstream workflows Pushes made by GITHUB_TOKEN don't trigger other workflows (GitHub Actions limitation to prevent infinite loops). This meant the auto-merge into int-agentic never triggered promote-to-main.yml. Now uses secrets.GH_PAT with fallback to GITHUB_TOKEN so: 1. If GH_PAT is configured: merge triggers promote-to-main.yml 2. If not: merge still works, but promotion PR must be created manually * Add workflow_dispatch trigger to promote-to-main Allows manual triggering when auto-trigger is missed (e.g. after GITHUB_TOKEN-based merges that don't fire downstream workflows). * Fix: Cosmos DB auth + remove phantom reflection_workflow_agent 1. MCP Cosmos DB: Replace AzureCliCredential (not available in containers) with ManagedIdentityCredential (when AZURE_CLIENT_ID is set) or DefaultAzureCredential as fallback. This fixes the 'Azure CLI not found on path' error in production Container Apps. 2. Backend: Remove reflection_workflow_agent from DEFAULT_AVAILABLE_AGENTS and descriptions - the module file doesn't exist, causing it to show as a broken option in the UI agent selector. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22321497108 * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Trigger rebuild: deploy Cosmos DB auth fix and agent list cleanup * Fix: iterate ResponseStream directly, not .updates ResponseStream is AsyncIterable - iterate with 'async for chunk in stream:' The .updates property returns a Sequence (list), not an async iterator. This caused TypeError: 'async for' requires __aiter__, got list. Not caught in local tests because unit tests mocked the streaming and the regression tests verified API signatures but didn't run live agents. * Fix: make integration test failures block the pipeline Removed continue-on-error: true from the pytest step so test failures actually fail the workflow. Previously, 5/7 tests could fail and the pipeline would still report success and proceed to auto-merge. * Fix: use SHA-tagged images for container deployment, not :latest Container Apps don't create a new revision when the image tag is unchanged. Using :latest meant the old container kept running even after a new image was pushed. Now passes github.sha as the image tag, which forces a new revision on every deploy. This was the root cause of the ResponseStream.updates bug reaching production - the integration tests ran against old containers. * Skip MCP localhost check in CI mode In CI, the evaluation talks to the backend via HTTP which connects to the deployed MCP container internally. The localhost:8000 check is only useful for local dev and was producing a confusing warning in CI logs. * Fix: auto-merge creates PR if none exists Previously auto-merge only looked for an existing open PR and skipped if none was found. Now it auto-creates a PR from the dev branch to int-agentic if one doesn't exist, then merges it. This ensures every successful pipeline run promotes code to int-agentic regardless of whether a PR was manually created beforehand. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22330447642 * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22334055353 --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * Bump agent-framework to 1.2.1 and migrate AzureOpenAIChatClient → OpenAIChatClient Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> * Rewrite handoff agent on native HandoffBuilder; update regression tests for 1.2.1 Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> * Update HANDOFF_README and dependency table for native HandoffBuilder migration Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> * Address GPT-5.5 review: fix CheckpointStorage protocol, reflection_agent kwargs, observability sample Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> * Fix PEP 8 spacing in mcp_agent_demo files (code review nits) Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> * Regenerate uv.lock files; pass workflow_name to 1.2.x checkpoint listing Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/91307ce7-2444-4b5f-b75f-2b736bc022c5 Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> --------- Co-authored-by: James Nguyen <janguy@microsoft.com> Co-authored-by: James N. <james.nguyen@microsoft.com> Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com> Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> --------- Co-authored-by: James N. <james.nguyen@microsoft.com> Co-authored-by: Copilot <198982749+Copilot@users.noreply.github.com> Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com>

…odel=…) for agent-framework 1.2.1 (#421) * Promote: int-agentic → main (production) (#414) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * Updating Durable Agent Implementation (#404) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * chore: reduce anomaly probability to 1% for controlled demo pace * Reduce anomaly probability to 1% for controlled demo pace (#406) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * chore: reduce anomaly probability to 1% for controlled demo pace --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * add mcp_agent_demo * Add hybrid MCP server (strict-schema + natural-language tools) and typed-contract workflow Scripts 6-8 for the MCP agent demo: - workflow_typed_contracts.py: 4-agent IT security incident response pipeline with Pydantic-enforced contracts at every boundary (SecurityAlert, ThreatAssessment, ImpactAnalysis, IncidentResponse) - mcp_server_hybrid.py: Single MCP endpoint exposing BOTH tool types: * Strict-schema: triage_alert, assess_threat, create_response (Pydantic) * Natural-language: ask_security_advisor, explain_for_customer (prose) * Shared session state across all tools - mcp_client_hybrid.py: 5-step incident flow using both tool types with explicit context passing between steps - README.md: Updated from 6 to 8 capabilities with Hybrid Mode architecture diagram and Quick Start sections 7-8 * Replace workflow_local_remote.py with simplified proxy agent (Script 3) - Delete workflow_local_remote.py (old Script 3 with extra LLM call) - Add workflow_proxy_agent.py: MCPProxyAgent (BaseAgent, no LLM) calls MCP tools directly via call_tool() zero local LLM overhead - Simplify MCPProxyAgent: extract common _call() method, module-level _last_text() helper, remove verbose static methods - Fix workflow_typed_contracts.py: add async with context managers - Update README: merge Script 9 into Script 3 position, update architecture diagram, Quick Start, How It Works sections * Clean up MCP demos: remove comparison commentary, drop typed-contracts - workflow_proxy_agent.py: remove 'traditional vs proxy' framing MCPProxyAgent is the standard approach for remote agent integration - Delete workflow_typed_contracts.py: not relevant to MCP integration - README: 7 scripts, renumber hybrid 6-7, remove typed-contract sections * Add LangGraph + MAF GroupChat cross-framework demo (Scripts 8-9) - mcp_server_langgraph.py: LangGraph ReAct agent exposed as MCP server on port 8003 with architecture tools (pattern eval, migration estimate, tech stack recommendation) - workflow_group_chat.py: MAF GroupChatBuilder orchestration with LLM planner routing between local BusinessStrategist (MAF) and remote TechnicalArchitect (LangGraph via MCP) - pyproject.toml: add langgraph, langchain-openai, langchain-core deps - README.md: update to 9 scripts, add cross-framework architecture diagram, Quick Start sections, and dependency table entries * MCP agent demo: rewrite README as MCP-vs-A2A thesis, delete proxy agent, finalize group chat with inline Planner * README: Mermaid diagrams, professional tone, remove LinkedIn teasing language * README: reframe around multi-framework interop problem, two design patterns (Agent-as-Tool + Agent Adapter), add conceptual architecture * README: fix Mermaid diagrams - use br tags instead of \n for line breaks * README: fix A2A comparison table - elicitation supported, structured schemas not * Upgrade agent-framework to 1.0.0rc1 and fastmcp to 3.0.2 Breaking changes migrated: - agent-framework 1.0.0rc1: ChatAgent->Agent, AgentThread->AgentSession, run_stream->run(stream=True), model->default_options(ChatOptions), WorkflowOutputEvent/AgentRunEvent/RequestInfoEvent->unified WorkflowEvent, MagenticBuilder now uses constructor kwargs, orchestrations moved to agent_framework_orchestrations package - fastmcp 3.0.2: removed mcp version pin (now pulled by fastmcp), no code changes needed (fully compatible) Files updated: - agentic_ai/agents/agent_framework/single_agent.py - agentic_ai/agents/agent_framework/multi_agent/handoff_multi_domain_agent.py - agentic_ai/agents/agent_framework/multi_agent/reflection_agent.py - agentic_ai/agents/agent_framework/multi_agent/magentic_group.py - agentic_ai/applications/pyproject.toml (agent-framework==1.0.0rc1) - agentic_ai/workflow/fraud_detection_durable/pyproject.toml (agent-framework==1.0.0rc1) - mcp/pyproject.toml (fastmcp==3.0.2) - tests/test_agent_framework_rc1_regression.py (51 regression tests, all passing) * Fix CI: regenerate requirements.txt with agent-framework-core==1.0.0rc1 The Docker build uses requirements.txt (not pyproject.toml/uv.lock), and it still had agent-framework-core==1.0.0b260130 pinned, which conflicts with agent-framework==1.0.0rc1 requiring agent-framework-core==1.0.0rc1. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22315210980 * Fix: use GH_PAT for auto-merge to trigger downstream workflows Pushes made by GITHUB_TOKEN don't trigger other workflows (GitHub Actions limitation to prevent infinite loops). This meant the auto-merge into int-agentic never triggered promote-to-main.yml. Now uses secrets.GH_PAT with fallback to GITHUB_TOKEN so: 1. If GH_PAT is configured: merge triggers promote-to-main.yml 2. If not: merge still works, but promotion PR must be created manually * Add workflow_dispatch trigger to promote-to-main Allows manual triggering when auto-trigger is missed (e.g. after GITHUB_TOKEN-based merges that don't fire downstream workflows). * Fix: Cosmos DB auth + remove phantom reflection_workflow_agent 1. MCP Cosmos DB: Replace AzureCliCredential (not available in containers) with ManagedIdentityCredential (when AZURE_CLIENT_ID is set) or DefaultAzureCredential as fallback. This fixes the 'Azure CLI not found on path' error in production Container Apps. 2. Backend: Remove reflection_workflow_agent from DEFAULT_AVAILABLE_AGENTS and descriptions - the module file doesn't exist, causing it to show as a broken option in the UI agent selector. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22321497108 * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Trigger rebuild: deploy Cosmos DB auth fix and agent list cleanup * Fix: iterate ResponseStream directly, not .updates ResponseStream is AsyncIterable - iterate with 'async for chunk in stream:' The .updates property returns a Sequence (list), not an async iterator. This caused TypeError: 'async for' requires __aiter__, got list. Not caught in local tests because unit tests mocked the streaming and the regression tests verified API signatures but didn't run live agents. * Fix: make integration test failures block the pipeline Removed continue-on-error: true from the pytest step so test failures actually fail the workflow. Previously, 5/7 tests could fail and the pipeline would still report success and proceed to auto-merge. * Fix: use SHA-tagged images for container deployment, not :latest Container Apps don't create a new revision when the image tag is unchanged. Using :latest meant the old container kept running even after a new image was pushed. Now passes github.sha as the image tag, which forces a new revision on every deploy. This was the root cause of the ResponseStream.updates bug reaching production - the integration tests ran against old containers. * Skip MCP localhost check in CI mode In CI, the evaluation talks to the backend via HTTP which connects to the deployed MCP container internally. The localhost:8000 check is only useful for local dev and was producing a confusing warning in CI logs. * Fix: auto-merge creates PR if none exists Previously auto-merge only looked for an existing open PR and skipped if none was found. Now it auto-creates a PR from the dev branch to int-agentic if one doesn't exist, then merges it. This ensures every successful pipeline run promotes code to int-agentic regardless of whether a PR was manually created beforehand. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22330447642 --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * Promote: int-agentic → main (production) (#416) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * Updating Durable Agent Implementation (#404) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * chore: reduce anomaly probability to 1% for controlled demo pace * Reduce anomaly probability to 1% for controlled demo pace (#406) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * chore: reduce anomaly probability to 1% for controlled demo pace --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * add mcp_agent_demo * Add hybrid MCP server (strict-schema + natural-language tools) and typed-contract workflow Scripts 6-8 for the MCP agent demo: - workflow_typed_contracts.py: 4-agent IT security incident response pipeline with Pydantic-enforced contracts at every boundary (SecurityAlert, ThreatAssessment, ImpactAnalysis, IncidentResponse) - mcp_server_hybrid.py: Single MCP endpoint exposing BOTH tool types: * Strict-schema: triage_alert, assess_threat, create_response (Pydantic) * Natural-language: ask_security_advisor, explain_for_customer (prose) * Shared session state across all tools - mcp_client_hybrid.py: 5-step incident flow using both tool types with explicit context passing between steps - README.md: Updated from 6 to 8 capabilities with Hybrid Mode architecture diagram and Quick Start sections 7-8 * Replace workflow_local_remote.py with simplified proxy agent (Script 3) - Delete workflow_local_remote.py (old Script 3 with extra LLM call) - Add workflow_proxy_agent.py: MCPProxyAgent (BaseAgent, no LLM) calls MCP tools directly via call_tool() zero local LLM overhead - Simplify MCPProxyAgent: extract common _call() method, module-level _last_text() helper, remove verbose static methods - Fix workflow_typed_contracts.py: add async with context managers - Update README: merge Script 9 into Script 3 position, update architecture diagram, Quick Start, How It Works sections * Clean up MCP demos: remove comparison commentary, drop typed-contracts - workflow_proxy_agent.py: remove 'traditional vs proxy' framing MCPProxyAgent is the standard approach for remote agent integration - Delete workflow_typed_contracts.py: not relevant to MCP integration - README: 7 scripts, renumber hybrid 6-7, remove typed-contract sections * Add LangGraph + MAF GroupChat cross-framework demo (Scripts 8-9) - mcp_server_langgraph.py: LangGraph ReAct agent exposed as MCP server on port 8003 with architecture tools (pattern eval, migration estimate, tech stack recommendation) - workflow_group_chat.py: MAF GroupChatBuilder orchestration with LLM planner routing between local BusinessStrategist (MAF) and remote TechnicalArchitect (LangGraph via MCP) - pyproject.toml: add langgraph, langchain-openai, langchain-core deps - README.md: update to 9 scripts, add cross-framework architecture diagram, Quick Start sections, and dependency table entries * MCP agent demo: rewrite README as MCP-vs-A2A thesis, delete proxy agent, finalize group chat with inline Planner * README: Mermaid diagrams, professional tone, remove LinkedIn teasing language * README: reframe around multi-framework interop problem, two design patterns (Agent-as-Tool + Agent Adapter), add conceptual architecture * README: fix Mermaid diagrams - use br tags instead of \n for line breaks * README: fix A2A comparison table - elicitation supported, structured schemas not * Upgrade agent-framework to 1.0.0rc1 and fastmcp to 3.0.2 Breaking changes migrated: - agent-framework 1.0.0rc1: ChatAgent->Agent, AgentThread->AgentSession, run_stream->run(stream=True), model->default_options(ChatOptions), WorkflowOutputEvent/AgentRunEvent/RequestInfoEvent->unified WorkflowEvent, MagenticBuilder now uses constructor kwargs, orchestrations moved to agent_framework_orchestrations package - fastmcp 3.0.2: removed mcp version pin (now pulled by fastmcp), no code changes needed (fully compatible) Files updated: - agentic_ai/agents/agent_framework/single_agent.py - agentic_ai/agents/agent_framework/multi_agent/handoff_multi_domain_agent.py - agentic_ai/agents/agent_framework/multi_agent/reflection_agent.py - agentic_ai/agents/agent_framework/multi_agent/magentic_group.py - agentic_ai/applications/pyproject.toml (agent-framework==1.0.0rc1) - agentic_ai/workflow/fraud_detection_durable/pyproject.toml (agent-framework==1.0.0rc1) - mcp/pyproject.toml (fastmcp==3.0.2) - tests/test_agent_framework_rc1_regression.py (51 regression tests, all passing) * Fix CI: regenerate requirements.txt with agent-framework-core==1.0.0rc1 The Docker build uses requirements.txt (not pyproject.toml/uv.lock), and it still had agent-framework-core==1.0.0b260130 pinned, which conflicts with agent-framework==1.0.0rc1 requiring agent-framework-core==1.0.0rc1. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22315210980 * Fix: use GH_PAT for auto-merge to trigger downstream workflows Pushes made by GITHUB_TOKEN don't trigger other workflows (GitHub Actions limitation to prevent infinite loops). This meant the auto-merge into int-agentic never triggered promote-to-main.yml. Now uses secrets.GH_PAT with fallback to GITHUB_TOKEN so: 1. If GH_PAT is configured: merge triggers promote-to-main.yml 2. If not: merge still works, but promotion PR must be created manually * Add workflow_dispatch trigger to promote-to-main Allows manual triggering when auto-trigger is missed (e.g. after GITHUB_TOKEN-based merges that don't fire downstream workflows). * Fix: Cosmos DB auth + remove phantom reflection_workflow_agent 1. MCP Cosmos DB: Replace AzureCliCredential (not available in containers) with ManagedIdentityCredential (when AZURE_CLIENT_ID is set) or DefaultAzureCredential as fallback. This fixes the 'Azure CLI not found on path' error in production Container Apps. 2. Backend: Remove reflection_workflow_agent from DEFAULT_AVAILABLE_AGENTS and descriptions - the module file doesn't exist, causing it to show as a broken option in the UI agent selector. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22321497108 * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Trigger rebuild: deploy Cosmos DB auth fix and agent list cleanup * Fix: iterate ResponseStream directly, not .updates ResponseStream is AsyncIterable - iterate with 'async for chunk in stream:' The .updates property returns a Sequence (list), not an async iterator. This caused TypeError: 'async for' requires __aiter__, got list. Not caught in local tests because unit tests mocked the streaming and the regression tests verified API signatures but didn't run live agents. * Fix: make integration test failures block the pipeline Removed continue-on-error: true from the pytest step so test failures actually fail the workflow. Previously, 5/7 tests could fail and the pipeline would still report success and proceed to auto-merge. * Fix: use SHA-tagged images for container deployment, not :latest Container Apps don't create a new revision when the image tag is unchanged. Using :latest meant the old container kept running even after a new image was pushed. Now passes github.sha as the image tag, which forces a new revision on every deploy. This was the root cause of the ResponseStream.updates bug reaching production - the integration tests ran against old containers. * Skip MCP localhost check in CI mode In CI, the evaluation talks to the backend via HTTP which connects to the deployed MCP container internally. The localhost:8000 check is only useful for local dev and was producing a confusing warning in CI logs. * Fix: auto-merge creates PR if none exists Previously auto-merge only looked for an existing open PR and skipped if none was found. Now it auto-creates a PR from the dev branch to int-agentic if one doesn't exist, then merges it. This ensures every successful pipeline run promotes code to int-agentic regardless of whether a PR was manually created beforehand. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22330447642 * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22334055353 --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * Promote: int-agentic → main (production) (#420) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * Updating Durable Agent Implementation (#404) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * chore: reduce anomaly probability to 1% for controlled demo pace * Reduce anomaly probability to 1% for controlled demo pace (#406) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * chore: reduce anomaly probability to 1% for controlled demo pace --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * add mcp_agent_demo * Add hybrid MCP server (strict-schema + natural-language tools) and typed-contract workflow Scripts 6-8 for the MCP agent demo: - workflow_typed_contracts.py: 4-agent IT security incident response pipeline with Pydantic-enforced contracts at every boundary (SecurityAlert, ThreatAssessment, ImpactAnalysis, IncidentResponse) - mcp_server_hybrid.py: Single MCP endpoint exposing BOTH tool types: * Strict-schema: triage_alert, assess_threat, create_response (Pydantic) * Natural-language: ask_security_advisor, explain_for_customer (prose) * Shared session state across all tools - mcp_client_hybrid.py: 5-step incident flow using both tool types with explicit context passing between steps - README.md: Updated from 6 to 8 capabilities with Hybrid Mode architecture diagram and Quick Start sections 7-8 * Replace workflow_local_remote.py with simplified proxy agent (Script 3) - Delete workflow_local_remote.py (old Script 3 with extra LLM call) - Add workflow_proxy_agent.py: MCPProxyAgent (BaseAgent, no LLM) calls MCP tools directly via call_tool() zero local LLM overhead - Simplify MCPProxyAgent: extract common _call() method, module-level _last_text() helper, remove verbose static methods - Fix workflow_typed_contracts.py: add async with context managers - Update README: merge Script 9 into Script 3 position, update architecture diagram, Quick Start, How It Works sections * Clean up MCP demos: remove comparison commentary, drop typed-contracts - workflow_proxy_agent.py: remove 'traditional vs proxy' framing MCPProxyAgent is the standard approach for remote agent integration - Delete workflow_typed_contracts.py: not relevant to MCP integration - README: 7 scripts, renumber hybrid 6-7, remove typed-contract sections * Add LangGraph + MAF GroupChat cross-framework demo (Scripts 8-9) - mcp_server_langgraph.py: LangGraph ReAct agent exposed as MCP server on port 8003 with architecture tools (pattern eval, migration estimate, tech stack recommendation) - workflow_group_chat.py: MAF GroupChatBuilder orchestration with LLM planner routing between local BusinessStrategist (MAF) and remote TechnicalArchitect (LangGraph via MCP) - pyproject.toml: add langgraph, langchain-openai, langchain-core deps - README.md: update to 9 scripts, add cross-framework architecture diagram, Quick Start sections, and dependency table entries * MCP agent demo: rewrite README as MCP-vs-A2A thesis, delete proxy agent, finalize group chat with inline Planner * README: Mermaid diagrams, professional tone, remove LinkedIn teasing language * README: reframe around multi-framework interop problem, two design patterns (Agent-as-Tool + Agent Adapter), add conceptual architecture * README: fix Mermaid diagrams - use br tags instead of \n for line breaks * README: fix A2A comparison table - elicitation supported, structured schemas not * Upgrade agent-framework to 1.0.0rc1 and fastmcp to 3.0.2 Breaking changes migrated: - agent-framework 1.0.0rc1: ChatAgent->Agent, AgentThread->AgentSession, run_stream->run(stream=True), model->default_options(ChatOptions), WorkflowOutputEvent/AgentRunEvent/RequestInfoEvent->unified WorkflowEvent, MagenticBuilder now uses constructor kwargs, orchestrations moved to agent_framework_orchestrations package - fastmcp 3.0.2: removed mcp version pin (now pulled by fastmcp), no code changes needed (fully compatible) Files updated: - agentic_ai/agents/agent_framework/single_agent.py - agentic_ai/agents/agent_framework/multi_agent/handoff_multi_domain_agent.py - agentic_ai/agents/agent_framework/multi_agent/reflection_agent.py - agentic_ai/agents/agent_framework/multi_agent/magentic_group.py - agentic_ai/applications/pyproject.toml (agent-framework==1.0.0rc1) - agentic_ai/workflow/fraud_detection_durable/pyproject.toml (agent-framework==1.0.0rc1) - mcp/pyproject.toml (fastmcp==3.0.2) - tests/test_agent_framework_rc1_regression.py (51 regression tests, all passing) * Fix CI: regenerate requirements.txt with agent-framework-core==1.0.0rc1 The Docker build uses requirements.txt (not pyproject.toml/uv.lock), and it still had agent-framework-core==1.0.0b260130 pinned, which conflicts with agent-framework==1.0.0rc1 requiring agent-framework-core==1.0.0rc1. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22315210980 * Fix: use GH_PAT for auto-merge to trigger downstream workflows Pushes made by GITHUB_TOKEN don't trigger other workflows (GitHub Actions limitation to prevent infinite loops). This meant the auto-merge into int-agentic never triggered promote-to-main.yml. Now uses secrets.GH_PAT with fallback to GITHUB_TOKEN so: 1. If GH_PAT is configured: merge triggers promote-to-main.yml 2. If not: merge still works, but promotion PR must be created manually * Add workflow_dispatch trigger to promote-to-main Allows manual triggering when auto-trigger is missed (e.g. after GITHUB_TOKEN-based merges that don't fire downstream workflows). * Fix: Cosmos DB auth + remove phantom reflection_workflow_agent 1. MCP Cosmos DB: Replace AzureCliCredential (not available in containers) with ManagedIdentityCredential (when AZURE_CLIENT_ID is set) or DefaultAzureCredential as fallback. This fixes the 'Azure CLI not found on path' error in production Container Apps. 2. Backend: Remove reflection_workflow_agent from DEFAULT_AVAILABLE_AGENTS and descriptions - the module file doesn't exist, causing it to show as a broken option in the UI agent selector. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22321497108 * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Trigger rebuild: deploy Cosmos DB auth fix and agent list cleanup * Fix: iterate ResponseStream directly, not .updates ResponseStream is AsyncIterable - iterate with 'async for chunk in stream:' The .updates property returns a Sequence (list), not an async iterator. This caused TypeError: 'async for' requires __aiter__, got list. Not caught in local tests because unit tests mocked the streaming and the regression tests verified API signatures but didn't run live agents. * Fix: make integration test failures block the pipeline Removed continue-on-error: true from the pytest step so test failures actually fail the workflow. Previously, 5/7 tests could fail and the pipeline would still report success and proceed to auto-merge. * Fix: use SHA-tagged images for container deployment, not :latest Container Apps don't create a new revision when the image tag is unchanged. Using :latest meant the old container kept running even after a new image was pushed. Now passes github.sha as the image tag, which forces a new revision on every deploy. This was the root cause of the ResponseStream.updates bug reaching production - the integration tests ran against old containers. * Skip MCP localhost check in CI mode In CI, the evaluation talks to the backend via HTTP which connects to the deployed MCP container internally. The localhost:8000 check is only useful for local dev and was producing a confusing warning in CI logs. * Fix: auto-merge creates PR if none exists Previously auto-merge only looked for an existing open PR and skipped if none was found. Now it auto-creates a PR from the dev branch to int-agentic if one doesn't exist, then merges it. This ensures every successful pipeline run promotes code to int-agentic regardless of whether a PR was manually created beforehand. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22330447642 * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22334055353 * Migrate workshop to agent-framework==1.2.1 (incl. native HandoffBuilder) (#419) * Promote: int-agentic → main (production) (#414) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * Updating Durable Agent Implementation (#404) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * chore: reduce anomaly probability to 1% for controlled demo pace * Reduce anomaly probability to 1% for controlled demo pace (#406) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * chore: reduce anomaly probability to 1% for controlled demo pace --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * add mcp_agent_demo * Add hybrid MCP server (strict-schema + natural-language tools) and typed-contract workflow Scripts 6-8 for the MCP agent demo: - workflow_typed_contracts.py: 4-agent IT security incident response pipeline with Pydantic-enforced contracts at every boundary (SecurityAlert, ThreatAssessment, ImpactAnalysis, IncidentResponse) - mcp_server_hybrid.py: Single MCP endpoint exposing BOTH tool types: * Strict-schema: triage_alert, assess_threat, create_response (Pydantic) * Natural-language: ask_security_advisor, explain_for_customer (prose) * Shared session state across all tools - mcp_client_hybrid.py: 5-step incident flow using both tool types with explicit context passing between steps - README.md: Updated from 6 to 8 capabilities with Hybrid Mode architecture diagram and Quick Start sections 7-8 * Replace workflow_local_remote.py with simplified proxy agent (Script 3) - Delete workflow_local_remote.py (old Script 3 with extra LLM call) - Add workflow_proxy_agent.py: MCPProxyAgent (BaseAgent, no LLM) calls MCP tools directly via call_tool() zero local LLM overhead - Simplify MCPProxyAgent: extract common _call() method, module-level _last_text() helper, remove verbose static methods - Fix workflow_typed_contracts.py: add async with context managers - Update README: merge Script 9 into Script 3 position, update architecture diagram, Quick Start, How It Works sections * Clean up MCP demos: remove comparison commentary, drop typed-contracts - workflow_proxy_agent.py: remove 'traditional vs proxy' framing MCPProxyAgent is the standard approach for remote agent integration - Delete workflow_typed_contracts.py: not relevant to MCP integration - README: 7 scripts, renumber hybrid 6-7, remove typed-contract sections * Add LangGraph + MAF GroupChat cross-framework demo (Scripts 8-9) - mcp_server_langgraph.py: LangGraph ReAct agent exposed as MCP server on port 8003 with architecture tools (pattern eval, migration estimate, tech stack recommendation) - workflow_group_chat.py: MAF GroupChatBuilder orchestration with LLM planner routing between local BusinessStrategist (MAF) and remote TechnicalArchitect (LangGraph via MCP) - pyproject.toml: add langgraph, langchain-openai, langchain-core deps - README.md: update to 9 scripts, add cross-framework architecture diagram, Quick Start sections, and dependency table entries * MCP agent demo: rewrite README as MCP-vs-A2A thesis, delete proxy agent, finalize group chat with inline Planner * README: Mermaid diagrams, professional tone, remove LinkedIn teasing language * README: reframe around multi-framework interop problem, two design patterns (Agent-as-Tool + Agent Adapter), add conceptual architecture * README: fix Mermaid diagrams - use br tags instead of \n for line breaks * README: fix A2A comparison table - elicitation supported, structured schemas not * Upgrade agent-framework to 1.0.0rc1 and fastmcp to 3.0.2 Breaking changes migrated: - agent-framework 1.0.0rc1: ChatAgent->Agent, AgentThread->AgentSession, run_stream->run(stream=True), model->default_options(ChatOptions), WorkflowOutputEvent/AgentRunEvent/RequestInfoEvent->unified WorkflowEvent, MagenticBuilder now uses constructor kwargs, orchestrations moved to agent_framework_orchestrations package - fastmcp 3.0.2: removed mcp version pin (now pulled by fastmcp), no code changes needed (fully compatible) Files updated: - agentic_ai/agents/agent_framework/single_agent.py - agentic_ai/agents/agent_framework/multi_agent/handoff_multi_domain_agent.py - agentic_ai/agents/agent_framework/multi_agent/reflection_agent.py - agentic_ai/agents/agent_framework/multi_agent/magentic_group.py - agentic_ai/applications/pyproject.toml (agent-framework==1.0.0rc1) - agentic_ai/workflow/fraud_detection_durable/pyproject.toml (agent-framework==1.0.0rc1) - mcp/pyproject.toml (fastmcp==3.0.2) - tests/test_agent_framework_rc1_regression.py (51 regression tests, all passing) * Fix CI: regenerate requirements.txt with agent-framework-core==1.0.0rc1 The Docker build uses requirements.txt (not pyproject.toml/uv.lock), and it still had agent-framework-core==1.0.0b260130 pinned, which conflicts with agent-framework==1.0.0rc1 requiring agent-framework-core==1.0.0rc1. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22315210980 * Fix: use GH_PAT for auto-merge to trigger downstream workflows Pushes made by GITHUB_TOKEN don't trigger other workflows (GitHub Actions limitation to prevent infinite loops). This meant the auto-merge into int-agentic never triggered promote-to-main.yml. Now uses secrets.GH_PAT with fallback to GITHUB_TOKEN so: 1. If GH_PAT is configured: merge triggers promote-to-main.yml 2. If not: merge still works, but promotion PR must be created manually * Add workflow_dispatch trigger to promote-to-main Allows manual triggering when auto-trigger is missed (e.g. after GITHUB_TOKEN-based merges that don't fire downstream workflows). * Fix: Cosmos DB auth + remove phantom reflection_workflow_agent 1. MCP Cosmos DB: Replace AzureCliCredential (not available in containers) with ManagedIdentityCredential (when AZURE_CLIENT_ID is set) or DefaultAzureCredential as fallback. This fixes the 'Azure CLI not found on path' error in production Container Apps. 2. Backend: Remove reflection_workflow_agent from DEFAULT_AVAILABLE_AGENTS and descriptions - the module file doesn't exist, causing it to show as a broken option in the UI agent selector. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22321497108 * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Trigger rebuild: deploy Cosmos DB auth fix and agent list cleanup * Fix: iterate ResponseStream directly, not .updates ResponseStream is AsyncIterable - iterate with 'async for chunk in stream:' The .updates property returns a Sequence (list), not an async iterator. This caused TypeError: 'async for' requires __aiter__, got list. Not caught in local tests because unit tests mocked the streaming and the regression tests verified API signatures but didn't run live agents. * Fix: make integration test failures block the pipeline Removed continue-on-error: true from the pytest step so test failures actually fail the workflow. Previously, 5/7 tests could fail and the pipeline would still report success and proceed to auto-merge. * Fix: use SHA-tagged images for container deployment, not :latest Container Apps don't create a new revision when the image tag is unchanged. Using :latest meant the old container kept running even after a new image was pushed. Now passes github.sha as the image tag, which forces a new revision on every deploy. This was the root cause of the ResponseStream.updates bug reaching production - the integration tests ran against old containers. * Skip MCP localhost check in CI mode In CI, the evaluation talks to the backend via HTTP which connects to the deployed MCP container internally. The localhost:8000 check is only useful for local dev and was producing a confusing warning in CI logs. * Fix: auto-merge creates PR if none exists Previously auto-merge only looked for an existing open PR and skipped if none was found. Now it auto-creates a PR from the dev branch to int-agentic if one doesn't exist, then merges it. This ensures every successful pipeline run promotes code to int-agentic regardless of whether a PR was manually created beforehand. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22330447642 --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * Promote: int-agentic → main (production) (#416) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * Updating Durable Agent Implementation (#404) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * chore: reduce anomaly probability to 1% for controlled demo pace * Reduce anomaly probability to 1% for controlled demo pace (#406) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * chore: reduce anomaly probability to 1% for controlled demo pace --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * add mcp_agent_demo * Add hybrid MCP server (strict-schema + natural-language tools) and typed-contract workflow Scripts 6-8 for the MCP agent demo: - workflow_typed_contracts.py: 4-agent IT security incident response pipeline with Pydantic-enforced contracts at every boundary (SecurityAlert, ThreatAssessment, ImpactAnalysis, IncidentResponse) - mcp_server_hybrid.py: Single MCP endpoint exposing BOTH tool types: * Strict-schema: triage_alert, assess_threat, create_response (Pydantic) * Natural-language: ask_security_advisor, explain_for_customer (prose) * Shared session state across all tools - mcp_client_hybrid.py: 5-step incident flow using both tool types with explicit context passing between steps - README.md: Updated from 6 to 8 capabilities with Hybrid Mode architecture diagram and Quick Start sections 7-8 * Replace workflow_local_remote.py with simplified proxy agent (Script 3) - Delete workflow_local_remote.py (old Script 3 with extra LLM call) - Add workflow_proxy_agent.py: MCPProxyAgent (BaseAgent, no LLM) calls MCP tools directly via call_tool() zero local LLM overhead - Simplify MCPProxyAgent: extract common _call() method, module-level _last_text() helper, remove verbose static methods - Fix workflow_typed_contracts.py: add async with context managers - Update README: merge Script 9 into Script 3 position, update architecture diagram, Quick Start, How It Works sections * Clean up MCP demos: remove comparison commentary, drop typed-contracts - workflow_proxy_agent.py: remove 'traditional vs proxy' framing MCPProxyAgent is the standard approach for remote agent integration - Delete workflow_typed_contracts.py: not relevant to MCP integration - README: 7 scripts, renumber hybrid 6-7, remove typed-contract sections * Add LangGraph + MAF GroupChat cross-framework demo (Scripts 8-9) - mcp_server_langgraph.py: LangGraph ReAct agent exposed as MCP server on port 8003 with architecture tools (pattern eval, migration estimate, tech stack recommendation) - workflow_group_chat.py: MAF GroupChatBuilder orchestration with LLM planner routing between local BusinessStrategist (MAF) and remote TechnicalArchitect (LangGraph via MCP) - pyproject.toml: add langgraph, langchain-openai, langchain-core deps - README.md: update to 9 scripts, add cross-framework architecture diagram, Quick Start sections, and dependency table entries * MCP agent demo: rewrite README as MCP-vs-A2A thesis, delete proxy agent, finalize group chat with inline Planner * README: Mermaid diagrams, professional tone, remove LinkedIn teasing language * README: reframe around multi-framework interop problem, two design patterns (Agent-as-Tool + Agent Adapter), add conceptual architecture * README: fix Mermaid diagrams - use br tags instead of \n for line breaks * README: fix A2A comparison table - elicitation supported, structured schemas not * Upgrade agent-framework to 1.0.0rc1 and fastmcp to 3.0.2 Breaking changes migrated: - agent-framework 1.0.0rc1: ChatAgent->Agent, AgentThread->AgentSession, run_stream->run(stream=True), model->default_options(ChatOptions), WorkflowOutputEvent/AgentRunEvent/RequestInfoEvent->unified WorkflowEvent, MagenticBuilder now uses constructor kwargs, orchestrations moved to agent_framework_orchestrations package - fastmcp 3.0.2: removed mcp version pin (now pulled by fastmcp), no code changes needed (fully compatible) Files updated: - agentic_ai/agents/agent_framework/single_agent.py - agentic_ai/agents/agent_framework/multi_agent/handoff_multi_domain_agent.py - agentic_ai/agents/agent_framework/multi_agent/reflection_agent.py - agentic_ai/agents/agent_framework/multi_agent/magentic_group.py - agentic_ai/applications/pyproject.toml (agent-framework==1.0.0rc1) - agentic_ai/workflow/fraud_detection_durable/pyproject.toml (agent-framework==1.0.0rc1) - mcp/pyproject.toml (fastmcp==3.0.2) - tests/test_agent_framework_rc1_regression.py (51 regression tests, all passing) * Fix CI: regenerate requirements.txt with agent-framework-core==1.0.0rc1 The Docker build uses requirements.txt (not pyproject.toml/uv.lock), and it still had agent-framework-core==1.0.0b260130 pinned, which conflicts with agent-framework==1.0.0rc1 requiring agent-framework-core==1.0.0rc1. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22315210980 * Fix: use GH_PAT for auto-merge to trigger downstream workflows Pushes made by GITHUB_TOKEN don't trigger other workflows (GitHub Actions limitation to prevent infinite loops). This meant the auto-merge into int-agentic never triggered promote-to-main.yml. Now uses secrets.GH_PAT with fallback to GITHUB_TOKEN so: 1. If GH_PAT is configured: merge triggers promote-to-main.yml 2. If not: merge still works, but promotion PR must be created manually * Add workflow_dispatch trigger to promote-to-main Allows manual triggering when auto-trigger is missed (e.g. after GITHUB_TOKEN-based merges that don't fire downstream workflows). * Fix: Cosmos DB auth + remove phantom reflection_workflow_agent 1. MCP Cosmos DB: Replace AzureCliCredential (not available in containers) with ManagedIdentityCredential (when AZURE_CLIENT_ID is set) or DefaultAzureCredential as fallback. This fixes the 'Azure CLI not found on path' error in production Container Apps. 2. Backend: Remove reflection_workflow_agent from DEFAULT_AVAILABLE_AGENTS and descriptions - the module file doesn't exist, causing it to show as a broken option in the UI agent selector. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22321497108 * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Trigger rebuild: deploy Cosmos DB auth fix and agent list cleanup * Fix: iterate ResponseStream directly, not .updates ResponseStream is AsyncIterable - iterate with 'async for chunk in stream:' The .updates property returns a Sequence (list), not an async iterator. This caused TypeError: 'async for' requires __aiter__, got list. Not caught in local tests because unit tests mocked the streaming and the regression tests verified API signatures but didn't run live agents. * Fix: make integration test failures block the pipeline Removed continue-on-error: true from the pytest step so test failures actually fail the workflow. Previously, 5/7 tests could fail and the pipeline would still report success and proceed to auto-merge. * Fix: use SHA-tagged images for container deployment, not :latest Container Apps don't create a new revision when the image tag is unchanged. Using :latest meant the old container kept running even after a new image was pushed. Now passes github.sha as the image tag, which forces a new revision on every deploy. This was the root cause of the ResponseStream.updates bug reaching production - the integration tests ran against old containers. * Skip MCP localhost check in CI mode In CI, the evaluation talks to the backend via HTTP which connects to the deployed MCP container internally. The localhost:8000 check is only useful for local dev and was producing a confusing warning in CI logs. * Fix: auto-merge creates PR if none exists Previously auto-merge only looked for an existing open PR and skipped if none was found. Now it auto-creates a PR from the dev branch to int-agentic if one doesn't exist, then merges it. This ensures every successful pipeline run promotes code to int-agentic regardless of whether a PR was manually created beforehand. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22330447642 * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22334055353 --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * Bump agent-framework to 1.2.1 and migrate AzureOpenAIChatClient → OpenAIChatClient Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> * Rewrite handoff agent on native HandoffBuilder; update regression tests for 1.2.1 Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> * Update HANDOFF_README and dependency table for native HandoffBuilder migration Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> * Address GPT-5.5 review: fix CheckpointStorage protocol, reflection_agent kwargs, observability sample Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> * Fix PEP 8 spacing in mcp_agent_demo files (code review nits) Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> * Regenerate uv.lock files; pass workflow_name to 1.2.x checkpoint listing Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/91307ce7-2444-4b5f-b75f-2b736bc022c5 Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> --------- Co-authored-by: James Nguyen <janguy@microsoft.com> Co-authored-by: James N. <james.nguyen@microsoft.com> Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com> Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> --------- Co-authored-by: James N. <james.nguyen@microsoft.com> Co-authored-by: Copilot <198982749+Copilot@users.noreply.github.com> Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> * Fix ChatOptions key for agent-framework 1.2.1 (model_id → model) Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/f921a448-fe80-4a39-8834-6b0ee615f1d1 Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> --------- Co-authored-by: James Nguyen <janguy@microsoft.com> Co-authored-by: James N. <james.nguyen@microsoft.com> Co-authored-by: Copilot <198982749+Copilot@users.noreply.github.com> Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> Co-authored-by: copilot <copilot@github.com>

* update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * Updating Durable Agent Implementation (#404) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * chore: reduce anomaly probability to 1% for controlled demo pace * Reduce anomaly probability to 1% for controlled demo pace (#406) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * chore: reduce anomaly probability to 1% for controlled demo pace --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * add mcp_agent_demo * Add hybrid MCP server (strict-schema + natural-language tools) and typed-contract workflow Scripts 6-8 for the MCP agent demo: - workflow_typed_contracts.py: 4-agent IT security incident response pipeline with Pydantic-enforced contracts at every boundary (SecurityAlert, ThreatAssessment, ImpactAnalysis, IncidentResponse) - mcp_server_hybrid.py: Single MCP endpoint exposing BOTH tool types: * Strict-schema: triage_alert, assess_threat, create_response (Pydantic) * Natural-language: ask_security_advisor, explain_for_customer (prose) * Shared session state across all tools - mcp_client_hybrid.py: 5-step incident flow using both tool types with explicit context passing between steps - README.md: Updated from 6 to 8 capabilities with Hybrid Mode architecture diagram and Quick Start sections 7-8 * Replace workflow_local_remote.py with simplified proxy agent (Script 3) - Delete workflow_local_remote.py (old Script 3 with extra LLM call) - Add workflow_proxy_agent.py: MCPProxyAgent (BaseAgent, no LLM) calls MCP tools directly via call_tool() zero local LLM overhead - Simplify MCPProxyAgent: extract common _call() method, module-level _last_text() helper, remove verbose static methods - Fix workflow_typed_contracts.py: add async with context managers - Update README: merge Script 9 into Script 3 position, update architecture diagram, Quick Start, How It Works sections * Clean up MCP demos: remove comparison commentary, drop typed-contracts - workflow_proxy_agent.py: remove 'traditional vs proxy' framing MCPProxyAgent is the standard approach for remote agent integration - Delete workflow_typed_contracts.py: not relevant to MCP integration - README: 7 scripts, renumber hybrid 6-7, remove typed-contract sections * Add LangGraph + MAF GroupChat cross-framework demo (Scripts 8-9) - mcp_server_langgraph.py: LangGraph ReAct agent exposed as MCP server on port 8003 with architecture tools (pattern eval, migration estimate, tech stack recommendation) - workflow_group_chat.py: MAF GroupChatBuilder orchestration with LLM planner routing between local BusinessStrategist (MAF) and remote TechnicalArchitect (LangGraph via MCP) - pyproject.toml: add langgraph, langchain-openai, langchain-core deps - README.md: update to 9 scripts, add cross-framework architecture diagram, Quick Start sections, and dependency table entries * MCP agent demo: rewrite README as MCP-vs-A2A thesis, delete proxy agent, finalize group chat with inline Planner * README: Mermaid diagrams, professional tone, remove LinkedIn teasing language * README: reframe around multi-framework interop problem, two design patterns (Agent-as-Tool + Agent Adapter), add conceptual architecture * README: fix Mermaid diagrams - use br tags instead of \n for line breaks * README: fix A2A comparison table - elicitation supported, structured schemas not * Upgrade agent-framework to 1.0.0rc1 and fastmcp to 3.0.2 Breaking changes migrated: - agent-framework 1.0.0rc1: ChatAgent->Agent, AgentThread->AgentSession, run_stream->run(stream=True), model->default_options(ChatOptions), WorkflowOutputEvent/AgentRunEvent/RequestInfoEvent->unified WorkflowEvent, MagenticBuilder now uses constructor kwargs, orchestrations moved to agent_framework_orchestrations package - fastmcp 3.0.2: removed mcp version pin (now pulled by fastmcp), no code changes needed (fully compatible) Files updated: - agentic_ai/agents/agent_framework/single_agent.py - agentic_ai/agents/agent_framework/multi_agent/handoff_multi_domain_agent.py - agentic_ai/agents/agent_framework/multi_agent/reflection_agent.py - agentic_ai/agents/agent_framework/multi_agent/magentic_group.py - agentic_ai/applications/pyproject.toml (agent-framework==1.0.0rc1) - agentic_ai/workflow/fraud_detection_durable/pyproject.toml (agent-framework==1.0.0rc1) - mcp/pyproject.toml (fastmcp==3.0.2) - tests/test_agent_framework_rc1_regression.py (51 regression tests, all passing) * Fix CI: regenerate requirements.txt with agent-framework-core==1.0.0rc1 The Docker build uses requirements.txt (not pyproject.toml/uv.lock), and it still had agent-framework-core==1.0.0b260130 pinned, which conflicts with agent-framework==1.0.0rc1 requiring agent-framework-core==1.0.0rc1. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22315210980 * Fix: use GH_PAT for auto-merge to trigger downstream workflows Pushes made by GITHUB_TOKEN don't trigger other workflows (GitHub Actions limitation to prevent infinite loops). This meant the auto-merge into int-agentic never triggered promote-to-main.yml. Now uses secrets.GH_PAT with fallback to GITHUB_TOKEN so: 1. If GH_PAT is configured: merge triggers promote-to-main.yml 2. If not: merge still works, but promotion PR must be created manually * Add workflow_dispatch trigger to promote-to-main Allows manual triggering when auto-trigger is missed (e.g. after GITHUB_TOKEN-based merges that don't fire downstream workflows). * Fix: Cosmos DB auth + remove phantom reflection_workflow_agent 1. MCP Cosmos DB: Replace AzureCliCredential (not available in containers) with ManagedIdentityCredential (when AZURE_CLIENT_ID is set) or DefaultAzureCredential as fallback. This fixes the 'Azure CLI not found on path' error in production Container Apps. 2. Backend: Remove reflection_workflow_agent from DEFAULT_AVAILABLE_AGENTS and descriptions - the module file doesn't exist, causing it to show as a broken option in the UI agent selector. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22321497108 * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Trigger rebuild: deploy Cosmos DB auth fix and agent list cleanup * Fix: iterate ResponseStream directly, not .updates ResponseStream is AsyncIterable - iterate with 'async for chunk in stream:' The .updates property returns a Sequence (list), not an async iterator. This caused TypeError: 'async for' requires __aiter__, got list. Not caught in local tests because unit tests mocked the streaming and the regression tests verified API signatures but didn't run live agents. * Fix: make integration test failures block the pipeline Removed continue-on-error: true from the pytest step so test failures actually fail the workflow. Previously, 5/7 tests could fail and the pipeline would still report success and proceed to auto-merge. * Fix: use SHA-tagged images for container deployment, not :latest Container Apps don't create a new revision when the image tag is unchanged. Using :latest meant the old container kept running even after a new image was pushed. Now passes github.sha as the image tag, which forces a new revision on every deploy. This was the root cause of the ResponseStream.updates bug reaching production - the integration tests ran against old containers. * Skip MCP localhost check in CI mode In CI, the evaluation talks to the backend via HTTP which connects to the deployed MCP container internally. The localhost:8000 check is only useful for local dev and was producing a confusing warning in CI logs. * Fix: auto-merge creates PR if none exists Previously auto-merge only looked for an existing open PR and skipped if none was found. Now it auto-creates a PR from the dev branch to int-agentic if one doesn't exist, then merges it. This ensures every successful pipeline run promotes code to int-agentic regardless of whether a PR was manually created beforehand. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22330447642 * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22334055353 * Migrate workshop to agent-framework==1.2.1 (incl. native HandoffBuilder) (#419) * Promote: int-agentic → main (production) (#414) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * Updating Durable Agent Implementation (#404) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * chore: reduce anomaly probability to 1% for controlled demo pace * Reduce anomaly probability to 1% for controlled demo pace (#406) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * chore: reduce anomaly probability to 1% for controlled demo pace --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * add mcp_agent_demo * Add hybrid MCP server (strict-schema + natural-language tools) and typed-contract workflow Scripts 6-8 for the MCP agent demo: - workflow_typed_contracts.py: 4-agent IT security incident response pipeline with Pydantic-enforced contracts at every boundary (SecurityAlert, ThreatAssessment, ImpactAnalysis, IncidentResponse) - mcp_server_hybrid.py: Single MCP endpoint exposing BOTH tool types: * Strict-schema: triage_alert, assess_threat, create_response (Pydantic) * Natural-language: ask_security_advisor, explain_for_customer (prose) * Shared session state across all tools - mcp_client_hybrid.py: 5-step incident flow using both tool types with explicit context passing between steps - README.md: Updated from 6 to 8 capabilities with Hybrid Mode architecture diagram and Quick Start sections 7-8 * Replace workflow_local_remote.py with simplified proxy agent (Script 3) - Delete workflow_local_remote.py (old Script 3 with extra LLM call) - Add workflow_proxy_agent.py: MCPProxyAgent (BaseAgent, no LLM) calls MCP tools directly via call_tool() zero local LLM overhead - Simplify MCPProxyAgent: extract common _call() method, module-level _last_text() helper, remove verbose static methods - Fix workflow_typed_contracts.py: add async with context managers - Update README: merge Script 9 into Script 3 position, update architecture diagram, Quick Start, How It Works sections * Clean up MCP demos: remove comparison commentary, drop typed-contracts - workflow_proxy_agent.py: remove 'traditional vs proxy' framing MCPProxyAgent is the standard approach for remote agent integration - Delete workflow_typed_contracts.py: not relevant to MCP integration - README: 7 scripts, renumber hybrid 6-7, remove typed-contract sections * Add LangGraph + MAF GroupChat cross-framework demo (Scripts 8-9) - mcp_server_langgraph.py: LangGraph ReAct agent exposed as MCP server on port 8003 with architecture tools (pattern eval, migration estimate, tech stack recommendation) - workflow_group_chat.py: MAF GroupChatBuilder orchestration with LLM planner routing between local BusinessStrategist (MAF) and remote TechnicalArchitect (LangGraph via MCP) - pyproject.toml: add langgraph, langchain-openai, langchain-core deps - README.md: update to 9 scripts, add cross-framework architecture diagram, Quick Start sections, and dependency table entries * MCP agent demo: rewrite README as MCP-vs-A2A thesis, delete proxy agent, finalize group chat with inline Planner * README: Mermaid diagrams, professional tone, remove LinkedIn teasing language * README: reframe around multi-framework interop problem, two design patterns (Agent-as-Tool + Agent Adapter), add conceptual architecture * README: fix Mermaid diagrams - use br tags instead of \n for line breaks * README: fix A2A comparison table - elicitation supported, structured schemas not * Upgrade agent-framework to 1.0.0rc1 and fastmcp to 3.0.2 Breaking changes migrated: - agent-framework 1.0.0rc1: ChatAgent->Agent, AgentThread->AgentSession, run_stream->run(stream=True), model->default_options(ChatOptions), WorkflowOutputEvent/AgentRunEvent/RequestInfoEvent->unified WorkflowEvent, MagenticBuilder now uses constructor kwargs, orchestrations moved to agent_framework_orchestrations package - fastmcp 3.0.2: removed mcp version pin (now pulled by fastmcp), no code changes needed (fully compatible) Files updated: - agentic_ai/agents/agent_framework/single_agent.py - agentic_ai/agents/agent_framework/multi_agent/handoff_multi_domain_agent.py - agentic_ai/agents/agent_framework/multi_agent/reflection_agent.py - agentic_ai/agents/agent_framework/multi_agent/magentic_group.py - agentic_ai/applications/pyproject.toml (agent-framework==1.0.0rc1) - agentic_ai/workflow/fraud_detection_durable/pyproject.toml (agent-framework==1.0.0rc1) - mcp/pyproject.toml (fastmcp==3.0.2) - tests/test_agent_framework_rc1_regression.py (51 regression tests, all passing) * Fix CI: regenerate requirements.txt with agent-framework-core==1.0.0rc1 The Docker build uses requirements.txt (not pyproject.toml/uv.lock), and it still had agent-framework-core==1.0.0b260130 pinned, which conflicts with agent-framework==1.0.0rc1 requiring agent-framework-core==1.0.0rc1. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22315210980 * Fix: use GH_PAT for auto-merge to trigger downstream workflows Pushes made by GITHUB_TOKEN don't trigger other workflows (GitHub Actions limitation to prevent infinite loops). This meant the auto-merge into int-agentic never triggered promote-to-main.yml. Now uses secrets.GH_PAT with fallback to GITHUB_TOKEN so: 1. If GH_PAT is configured: merge triggers promote-to-main.yml 2. If not: merge still works, but promotion PR must be created manually * Add workflow_dispatch trigger to promote-to-main Allows manual triggering when auto-trigger is missed (e.g. after GITHUB_TOKEN-based merges that don't fire downstream workflows). * Fix: Cosmos DB auth + remove phantom reflection_workflow_agent 1. MCP Cosmos DB: Replace AzureCliCredential (not available in containers) with ManagedIdentityCredential (when AZURE_CLIENT_ID is set) or DefaultAzureCredential as fallback. This fixes the 'Azure CLI not found on path' error in production Container Apps. 2. Backend: Remove reflection_workflow_agent from DEFAULT_AVAILABLE_AGENTS and descriptions - the module file doesn't exist, causing it to show as a broken option in the UI agent selector. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22321497108 * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Trigger rebuild: deploy Cosmos DB auth fix and agent list cleanup * Fix: iterate ResponseStream directly, not .updates ResponseStream is AsyncIterable - iterate with 'async for chunk in stream:' The .updates property returns a Sequence (list), not an async iterator. This caused TypeError: 'async for' requires __aiter__, got list. Not caught in local tests because unit tests mocked the streaming and the regression tests verified API signatures but didn't run live agents. * Fix: make integration test failures block the pipeline Removed continue-on-error: true from the pytest step so test failures actually fail the workflow. Previously, 5/7 tests could fail and the pipeline would still report success and proceed to auto-merge. * Fix: use SHA-tagged images for container deployment, not :latest Container Apps don't create a new revision when the image tag is unchanged. Using :latest meant the old container kept running even after a new image was pushed. Now passes github.sha as the image tag, which forces a new revision on every deploy. This was the root cause of the ResponseStream.updates bug reaching production - the integration tests ran against old containers. * Skip MCP localhost check in CI mode In CI, the evaluation talks to the backend via HTTP which connects to the deployed MCP container internally. The localhost:8000 check is only useful for local dev and was producing a confusing warning in CI logs. * Fix: auto-merge creates PR if none exists Previously auto-merge only looked for an existing open PR and skipped if none was found. Now it auto-creates a PR from the dev branch to int-agentic if one doesn't exist, then merges it. This ensures every successful pipeline run promotes code to int-agentic regardless of whether a PR was manually created beforehand. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22330447642 --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * Promote: int-agentic → main (production) (#416) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * Updating Durable Agent Implementation (#404) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * chore: reduce anomaly probability to 1% for controlled demo pace * Reduce anomaly probability to 1% for controlled demo pace (#406) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * chore: reduce anomaly probability to 1% for controlled demo pace --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * add mcp_agent_demo * Add hybrid MCP server (strict-schema + natural-language tools) and typed-contract workflow Scripts 6-8 for the MCP agent demo: - workflow_typed_contracts.py: 4-agent IT security incident response pipeline with Pydantic-enforced contracts at every boundary (SecurityAlert, ThreatAssessment, ImpactAnalysis, IncidentResponse) - mcp_server_hybrid.py: Single MCP endpoint exposing BOTH tool types: * Strict-schema: triage_alert, assess_threat, create_response (Pydantic) * Natural-language: ask_security_advisor, explain_for_customer (prose) * Shared session state across all tools - mcp_client_hybrid.py: 5-step incident flow using both tool types with explicit context passing between steps - README.md: Updated from 6 to 8 capabilities with Hybrid Mode architecture diagram and Quick Start sections 7-8 * Replace workflow_local_remote.py with simplified proxy agent (Script 3) - Delete workflow_local_remote.py (old Script 3 with extra LLM call) - Add workflow_proxy_agent.py: MCPProxyAgent (BaseAgent, no LLM) calls MCP tools directly via call_tool() zero local LLM overhead - Simplify MCPProxyAgent: extract common _call() method, module-level _last_text() helper, remove verbose static methods - Fix workflow_typed_contracts.py: add async with context managers - Update README: merge Script 9 into Script 3 position, update architecture diagram, Quick Start, How It Works sections * Clean up MCP demos: remove comparison commentary, drop typed-contracts - workflow_proxy_agent.py: remove 'traditional vs proxy' framing MCPProxyAgent is the standard approach for remote agent integration - Delete workflow_typed_contracts.py: not relevant to MCP integration - README: 7 scripts, renumber hybrid 6-7, remove typed-contract sections * Add LangGraph + MAF GroupChat cross-framework demo (Scripts 8-9) - mcp_server_langgraph.py: LangGraph ReAct agent exposed as MCP server on port 8003 with architecture tools (pattern eval, migration estimate, tech stack recommendation) - workflow_group_chat.py: MAF GroupChatBuilder orchestration with LLM planner routing between local BusinessStrategist (MAF) and remote TechnicalArchitect (LangGraph via MCP) - pyproject.toml: add langgraph, langchain-openai, langchain-core deps - README.md: update to 9 scripts, add cross-framework architecture diagram, Quick Start sections, and dependency table entries * MCP agent demo: rewrite README as MCP-vs-A2A thesis, delete proxy agent, finalize group chat with inline Planner * README: Mermaid diagrams, professional tone, remove LinkedIn teasing language * README: reframe around multi-framework interop problem, two design patterns (Agent-as-Tool + Agent Adapter), add conceptual architecture * README: fix Mermaid diagrams - use br tags instead of \n for line breaks * README: fix A2A comparison table - elicitation supported, structured schemas not * Upgrade agent-framework to 1.0.0rc1 and fastmcp to 3.0.2 Breaking changes migrated: - agent-framework 1.0.0rc1: ChatAgent->Agent, AgentThread->AgentSession, run_stream->run(stream=True), model->default_options(ChatOptions), WorkflowOutputEvent/AgentRunEvent/RequestInfoEvent->unified WorkflowEvent, MagenticBuilder now uses constructor kwargs, orchestrations moved to agent_framework_orchestrations package - fastmcp 3.0.2: removed mcp version pin (now pulled by fastmcp), no code changes needed (fully compatible) Files updated: - agentic_ai/agents/agent_framework/single_agent.py - agentic_ai/agents/agent_framework/multi_agent/handoff_multi_domain_agent.py - agentic_ai/agents/agent_framework/multi_agent/reflection_agent.py - agentic_ai/agents/agent_framework/multi_agent/magentic_group.py - agentic_ai/applications/pyproject.toml (agent-framework==1.0.0rc1) - agentic_ai/workflow/fraud_detection_durable/pyproject.toml (agent-framework==1.0.0rc1) - mcp/pyproject.toml (fastmcp==3.0.2) - tests/test_agent_framework_rc1_regression.py (51 regression tests, all passing) * Fix CI: regenerate requirements.txt with agent-framework-core==1.0.0rc1 The Docker build uses requirements.txt (not pyproject.toml/uv.lock), and it still had agent-framework-core==1.0.0b260130 pinned, which conflicts with agent-framework==1.0.0rc1 requiring agent-framework-core==1.0.0rc1. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22315210980 * Fix: use GH_PAT for auto-merge to trigger downstream workflows Pushes made by GITHUB_TOKEN don't trigger other workflows (GitHub Actions limitation to prevent infinite loops). This meant the auto-merge into int-agentic never triggered promote-to-main.yml. Now uses secrets.GH_PAT with fallback to GITHUB_TOKEN so: 1. If GH_PAT is configured: merge triggers promote-to-main.yml 2. If not: merge still works, but promotion PR must be created manually * Add workflow_dispatch trigger to promote-to-main Allows manual triggering when auto-trigger is missed (e.g. after GITHUB_TOKEN-based merges that don't fire downstream workflows). * Fix: Cosmos DB auth + remove phantom reflection_workflow_agent 1. MCP Cosmos DB: Replace AzureCliCredential (not available in containers) with ManagedIdentityCredential (when AZURE_CLIENT_ID is set) or DefaultAzureCredential as fallback. This fixes the 'Azure CLI not found on path' error in production Container Apps. 2. Backend: Remove reflection_workflow_agent from DEFAULT_AVAILABLE_AGENTS and descriptions - the module file doesn't exist, causing it to show as a broken option in the UI agent selector. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22321497108 * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Trigger rebuild: deploy Cosmos DB auth fix and agent list cleanup * Fix: iterate ResponseStream directly, not .updates ResponseStream is AsyncIterable - iterate with 'async for chunk in stream:' The .updates property returns a Sequence (list), not an async iterator. This caused TypeError: 'async for' requires __aiter__, got list. Not caught in local tests because unit tests mocked the streaming and the regression tests verified API signatures but didn't run live agents. * Fix: make integration test failures block the pipeline Removed continue-on-error: true from the pytest step so test failures actually fail the workflow. Previously, 5/7 tests could fail and the pipeline would still report success and proceed to auto-merge. * Fix: use SHA-tagged images for container deployment, not :latest Container Apps don't create a new revision when the image tag is unchanged. Using :latest meant the old container kept running even after a new image was pushed. Now passes github.sha as the image tag, which forces a new revision on every deploy. This was the root cause of the ResponseStream.updates bug reaching production - the integration tests ran against old containers. * Skip MCP localhost check in CI mode In CI, the evaluation talks to the backend via HTTP which connects to the deployed MCP container internally. The localhost:8000 check is only useful for local dev and was producing a confusing warning in CI logs. * Fix: auto-merge creates PR if none exists Previously auto-merge only looked for an existing open PR and skipped if none was found. Now it auto-creates a PR from the dev branch to int-agentic if one doesn't exist, then merges it. This ensures every successful pipeline run promotes code to int-agentic regardless of whether a PR was manually created beforehand. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22330447642 * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22334055353 --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * Bump agent-framework to 1.2.1 and migrate AzureOpenAIChatClient → OpenAIChatClient Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> * Rewrite handoff agent on native HandoffBuilder; update regression tests for 1.2.1 Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> * Update HANDOFF_README and dependency table for native HandoffBuilder migration Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> * Address GPT-5.5 review: fix CheckpointStorage protocol, reflection_agent kwargs, observability sample Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> * Fix PEP 8 spacing in mcp_agent_demo files (code review nits) Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> * Regenerate uv.lock files; pass workflow_name to 1.2.x checkpoint listing Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/91307ce7-2444-4b5f-b75f-2b736bc022c5 Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> --------- Co-authored-by: James Nguyen <janguy@microsoft.com> Co-authored-by: James N. <james.nguyen@microsoft.com> Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com> Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> * docs: Remove misleading RequestInfoExecutor references (#418) doc update * Fix backend 500s by renaming ChatOptions(model_id=…) to ChatOptions(model=…) for agent-framework 1.2.1 (#421) * Promote: int-agentic → main (production) (#414) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * Updating Durable Agent Implementation (#404) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * chore: reduce anomaly probability to 1% for controlled demo pace * Reduce anomaly probability to 1% for controlled demo pace (#406) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * chore: reduce anomaly probability to 1% for controlled demo pace --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * add mcp_agent_demo * Add hybrid MCP server (strict-schema + natural-language tools) and typed-contract workflow Scripts 6-8 for the MCP agent demo: - workflow_typed_contracts.py: 4-agent IT security incident response pipeline with Pydantic-enforced contracts at every boundary (SecurityAlert, ThreatAssessment, ImpactAnalysis, IncidentResponse) - mcp_server_hybrid.py: Single MCP endpoint exposing BOTH tool types: * Strict-schema: triage_alert, assess_threat, create_response (Pydantic) * Natural-language: ask_security_advisor, explain_for_customer (prose) * Shared session state across all tools - mcp_client_hybrid.py: 5-step incident flow using both tool types with explicit context passing between steps - README.md: Updated from 6 to 8 capabilities with Hybrid Mode architecture diagram and Quick Start sections 7-8 * Replace workflow_local_remote.py with simplified proxy agent (Script 3) - Delete workflow_local_remote.py (old Script 3 with extra LLM call) - Add workflow_proxy_agent.py: MCPProxyAgent (BaseAgent, no LLM) calls MCP tools directly via call_tool() zero local LLM overhead - Simplify MCPProxyAgent: extract common _call() method, module-level _last_text() helper, remove verbose static methods - Fix workflow_typed_contracts.py: add async with context managers - Update README: merge Script 9 into Script 3 position, update architecture diagram, Quick Start, How It Works sections * Clean up MCP demos: remove comparison commentary, drop typed-contracts - workflow_proxy_agent.py: remove 'traditional vs proxy' framing MCPProxyAgent is the standard approach for remote agent integration - Delete workflow_typed_contracts.py: not relevant to MCP integration - README: 7 scripts, renumber hybrid 6-7, remove typed-contract sections * Add LangGraph + MAF GroupChat cross-framework demo (Scripts 8-9) - mcp_server_langgraph.py: LangGraph ReAct agent exposed as MCP server on port 8003 with architecture tools (pattern eval, migration estimate, tech stack recommendation) - workflow_group_chat.py: MAF GroupChatBuilder orchestration with LLM planner routing between local BusinessStrategist (MAF) and remote TechnicalArchitect (LangGraph via MCP) - pyproject.toml: add langgraph, langchain-openai, langchain-core deps - README.md: update to 9 scripts, add cross-framework architecture diagram, Quick Start sections, and dependency table entries * MCP agent demo: rewrite README as MCP-vs-A2A thesis, delete proxy agent, finalize group chat with inline Planner * README: Mermaid diagrams, professional tone, remove LinkedIn teasing language * README: reframe around multi-framework interop problem, two design patterns (Agent-as-Tool + Agent Adapter), add conceptual architecture * README: fix Mermaid diagrams - use br tags instead of \n for line breaks * README: fix A2A comparison table - elicitation supported, structured schemas not * Upgrade agent-framework to 1.0.0rc1 and fastmcp to 3.0.2 Breaking changes migrated: - agent-framework 1.0.0rc1: ChatAgent->Agent, AgentThread->AgentSession, run_stream->run(stream=True), model->default_options(ChatOptions), WorkflowOutputEvent/AgentRunEvent/RequestInfoEvent->unified WorkflowEvent, MagenticBuilder now uses constructor kwargs, orchestrations moved to agent_framework_orchestrations package - fastmcp 3.0.2: removed mcp version pin (now pulled by fastmcp), no code changes needed (fully compatible) Files updated: - agentic_ai/agents/agent_framework/single_agent.py - agentic_ai/agents/agent_framework/multi_agent/handoff_multi_domain_agent.py - agentic_ai/agents/agent_framework/multi_agent/reflection_agent.py - agentic_ai/agents/agent_framework/multi_agent/magentic_group.py - agentic_ai/applications/pyproject.toml (agent-framework==1.0.0rc1) - agentic_ai/workflow/fraud_detection_durable/pyproject.toml (agent-framework==1.0.0rc1) - mcp/pyproject.toml (fastmcp==3.0.2) - tests/test_agent_framework_rc1_regression.py (51 regression tests, all passing) * Fix CI: regenerate requirements.txt with agent-framework-core==1.0.0rc1 The Docker build uses requirements.txt (not pyproject.toml/uv.lock), and it still had agent-framework-core==1.0.0b260130 pinned, which conflicts with agent-framework==1.0.0rc1 requiring agent-framework-core==1.0.0rc1. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22315210980 * Fix: use GH_PAT for auto-merge to trigger downstream workflows Pushes made by GITHUB_TOKEN don't trigger other workflows (GitHub Actions limitation to prevent infinite loops). This meant the auto-merge into int-agentic never triggered promote-to-main.yml. Now uses secrets.GH_PAT with fallback to GITHUB_TOKEN so: 1. If GH_PAT is configured: merge triggers promote-to-main.yml 2. If not: merge still works, but promotion PR must be created manually * Add workflow_dispatch trigger to promote-to-main Allows manual triggering when auto-trigger is missed (e.g. after GITHUB_TOKEN-based merges that don't fire downstream workflows). * Fix: Cosmos DB auth + remove phantom reflection_workflow_agent 1. MCP Cosmos DB: Replace AzureCliCredential (not available in containers) with ManagedIdentityCredential (when AZURE_CLIENT_ID is set) or DefaultAzureCredential as fallback. This fixes the 'Azure CLI not found on path' error in production Container Apps. 2. Backend: Remove reflection_workflow_agent from DEFAULT_AVAILABLE_AGENTS and descriptions - the module file doesn't exist, causing it to show as a broken option in the UI agent selector. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22321497108 * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Trigger rebuild: deploy Cosmos DB auth fix and agent list cleanup * Fix: iterate ResponseStream directly, not .updates ResponseStream is AsyncIterable - iterate with 'async for chunk in stream:' The .updates property returns a Sequence (list), not an async iterator. This caused TypeError: 'async for' requires __aiter__, got list. Not caught in local tests because unit tests mocked the streaming and the regression tests verified API signatures but didn't run live agents. * Fix: make integration test failures block the pipeline Removed continue-on-error: true from the pytest step so test failures actually fail the workflow. Previously, 5/7 tests could fail and the pipeline would still report success and proceed to auto-merge. * Fix: use SHA-tagged images for container deployment, not :latest Container Apps don't create a new revision when the image tag is unchanged. Using :latest meant the old container kept running even after a new image was pushed. Now passes github.sha as the image tag, which forces a new revision on every deploy. This was the root cause of the ResponseStream.updates bug reaching production - the integration tests ran against old containers. * Skip MCP localhost check in CI mode In CI, the evaluation talks to the backend via HTTP which connects to the deployed MCP container internally. The localhost:8000 check is only useful for local dev and was producing a confusing warning in CI logs. * Fix: auto-merge creates PR if none exists Previously auto-merge only looked for an existing open PR and skipped if none was found. Now it auto-creates a PR from the dev branch to int-agentic if one doesn't exist, then merges it. This ensures every successful pipeline run promotes code to int-agentic regardless of whether a PR was manually created beforehand. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22330447642 --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * Promote: int-agentic → main (production) (#416) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * Updating Durable Agent Implementation (#404) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * chore: reduce anomaly probability to 1% for controlled demo pace * Reduce anomaly probability to 1% for controlled demo pace (#406) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * chore: reduce anomaly probability to 1% for controlled demo pace --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * add mcp_agent_demo * Add hybrid MCP server (strict-schema + natural-language tools) and typed-contract workflow Scripts 6-8 for the MCP agent demo: - workflow_typed_contracts.py: 4-agent IT security incident response pipeline with Pydantic-enforced contracts at every boundary (SecurityAlert, ThreatAssessment, ImpactAnalysis, IncidentResponse) - mcp_server_hybrid.py: Single MCP endpoint exposing BOTH tool types: * Strict-schema: triage_alert, assess_threat, create_response (Pydantic) * Natural-language: ask_security_advisor, explain_for_customer (prose) * Shared session state across all tools - mcp_client_hybrid.py: 5-step incident flow using both tool types with explicit context passing between steps - README.md: Updated from 6 to 8 capabilities with Hybrid Mode architecture diagram and Quick Start sections 7-8 * Replace workflow_local_remote.py with simplified proxy agent (Script 3) - Delete workflow_local_remote.py (old Script 3 with extra LLM call) - Add workflow_proxy_agent.py: MCPProxyAgent (BaseAgent, no LLM) calls MCP tools directly via call_tool() zero local LLM overhead - Simplify MCPProxyAgent: extract common _call() method, module-level _last_text() helper, remove verbose static methods - Fix workflow_typed_contracts.py: add async with context managers - Update README: merge Script 9 into Script 3 position, update architecture diagram, Quick Start, How It Works sections * Clean up MCP demos: remove comparison commentary, drop typed-contracts - workflow_proxy_agent.py: remove 'traditional vs proxy' framing MCPProxyAgent is the standard approach for remote agent integration - Delete workflow_typed_contracts.py: not relevant to MCP integration - README: 7 scripts, renumber hybrid 6-7, remove typed-contract sections * Add LangGraph + MAF GroupChat cross-framework demo (Scripts 8-9) - mcp_server_langgraph.py: LangGraph ReAct agent exposed as MCP server on port 8003 with architecture tools (pattern eval, migration estimate, tech stack recommendation) - workflow_group_chat.py: MAF GroupChatBuilder orchestration with LLM planner routing between local BusinessStrategist (MAF) and remote TechnicalArchitect (LangGraph via MCP) - pyproject.toml: add langgraph, langchain-openai, langchain-core deps - README.md: update to 9 scripts, add cross-framework architecture diagram, Quick Start sections, and dependency table entries * MCP agent demo: rewrite README as MCP-vs-A2A thesis, delete proxy agent, finalize group chat with inline Planner * README: Mermaid diagrams, professional tone, remove LinkedIn teasing language * README: reframe around multi-framework interop problem, two design patterns (Agent-as-Tool + Agent Adapter), add conceptual architecture * README: fix Mermaid diagrams - use br tags instead of \n for line breaks * README: fix A2A comparison table - elicitation supported, structured schemas not * Upgrade agent-framework to 1.0.0rc1 and fastmcp to 3.0.2 Breaking changes migrated: - agent-framework 1.0.0rc1: ChatAgent->Agent, AgentThread->AgentSession, run_stream->run(stream=True), model->default_options(ChatOptions), WorkflowOutputEvent/AgentRunEvent/RequestInfoEvent->unified WorkflowEvent, MagenticBuilder now uses constructor kwargs, orchestrations moved to agent_framework_orchestrations package - fastmcp 3.0.2: removed mcp version pin (now pulled by fastmcp), no code changes needed (fully compatible) Files updated: - agentic_ai/agents/agent_framework/single_agent.py - agentic_ai/agents/agent_framework/multi_agent/handoff_multi_domain_agent.py - agentic_ai/agents/agent_framework/multi_agent/reflection_agent.py - agentic_ai/agents/agent_framework/multi_agent/magentic_group.py - agentic_ai/applications/pyproject.toml (agent-framework==1.0.0rc1) - agentic_ai/workflow/fraud_detection_durable/pyproject.toml (agent-framework==1.0.0rc1) - mcp/pyproject.toml (fastmcp==3.0.2) - tests/test_agent_framework_rc1_regression.py (51 regression tests, all passing) * Fix CI: regenerate requirements.txt with agent-framework-core==1.0.0rc1 The Docker build uses requirements.txt (not pyproject.toml/uv.lock), and it still had agent-framework-core==1.0.0b260130 pinned, which conflicts with agent-framework==1.0.0rc1 requiring agent-framework-core==1.0.0rc1. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22315210980 * Fix: use GH_PAT for auto-merge to trigger downstream workflows Pushes made by GITHUB_TOKEN don't trigger other workflows (GitHub Actions limitation to prevent infinite loops). This meant the auto-merge into int-agentic never triggered promote-to-main.yml. Now uses secrets.GH_PAT with fallback to GITHUB_TOKEN so: 1. If GH_PAT is configured: merge triggers promote-to-main.yml 2. If not: merge still works, but promotion PR must be created manually * Add workflow_dispatch trigger to promote-to-main Allows manual triggering when auto-trigger is missed (e.g. after GITHUB_TOKEN-based merges that don't fire downstream workflows). * Fix: Cosmos DB auth + remove phantom reflection_workflow_agent 1. MCP Cosmos DB: Replace AzureCliCredential (not available in containers) with ManagedIdentityCredential (when AZURE_CLIENT_ID is set) or DefaultAzureCredential as fallback. This fixes the 'Azure CLI not found on path' error in production Container Apps. 2. Backend: Remove reflection_workflow_agent from DEFAULT_AVAILABLE_AGENTS and descriptions - the module file doesn't exist, causing it to show as a broken option in the UI agent selector. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22321497108 * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Trigger rebuild: deploy Cosmos DB auth fix and agent list cleanup * Fix: iterate ResponseStream directly, not .updates ResponseStream is AsyncIterable - iterate with 'async for chunk in stream:' The .updates property returns a Sequence (list), not an async iterator. This caused TypeError: 'async for' requires __aiter__, got list. Not caught in local tests because unit tests mocked the streaming and the regression tests verified API signatures but didn't run live agents. * Fix: make integration test failures block the pipeline Removed continue-on-error: true from the pytest step so test failures actually fail the workflow. Previously, 5/7 tests could fail and the pipeline would still report success and proceed to auto-merge. * Fix: use SHA-tagged images for container deployment, not :latest Container Apps don't create a new revision when the image tag is unchanged. Using :latest meant the old container kept running even after a new image was pushed. Now passes github.sha as the image tag, which forces a new revision on every deploy. This was the root cause of the ResponseStream.updates bug reaching production - the integration tests ran against old containers. * Skip MCP localhost check in CI mode In CI, the evaluation talks to the backend via HTTP which connects to the deployed MCP container internally. The localhost:8000 check is only useful for local dev and was producing a confusing warning in CI logs. * Fix: auto-merge creates PR if none exists Previously auto-merge only looked for an existing open PR and skipped if none was found. Now it auto-creates a PR from the dev branch to int-agentic if one doesn't exist, then merges it. This ensures every successful pipeline run promotes code to int-agentic regardless of whether a PR was manually created beforehand. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22330447642 * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22334055353 --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * Promote: int-agentic → main (production) (#420) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * Updating Durable Agent Implementation (#404) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * chore: reduce anomaly probability to 1% for controlled demo pace * Reduce anomaly probability to 1% for controlled demo pace (#406) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * chore: reduce anomaly probability to 1% for controlled demo pace --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * add mcp_agent_demo * Add hybrid MCP server (strict-schema + natural-language tools) and typed-contract workflow Scripts 6-8 for the MCP agent demo: - workflow_typed_contracts.py: 4-agent IT security incident response pipeline with Pydantic-enforced contracts at every boundary (SecurityAlert, ThreatAssessment, ImpactAnalysis, IncidentResponse) - mcp_server_hybrid.py: Single MCP endpoint exposing BOTH tool types: * Strict-schema: triage_alert, assess_threat, create_response (Pydantic) * Natural-language: ask_security_advisor, explain_for_customer (prose) * Shared session state across all tools - mcp_client_hybrid.py: 5-step incident flow using both tool types with explicit context passing between steps - README.md: Updated from 6 to 8 capabilities with Hybrid Mode architecture diagram and Quick Start sections 7-8 * Replace workflow_local_remote.py with simplified proxy agent (Script 3) - Delete workflow_local_remote.py (old Script 3 with extra LLM call) - Add workflow_proxy_agent.py: MCPProxyAgent (BaseAgent, no LLM) calls MCP tools directly via call_tool() zero local LLM overhead - Simplify MCPProxyAgent: extract common _call() method, module-level _last_text() helper, remove verbose static methods - Fix workflow_typed_contracts.py: add async with context managers - Update README: merge Script 9 into Script 3 position, update architecture diagram, Quick Start, How It Works sections * Clean up MCP demos: remove comparison commentary, drop typed-contracts - workflow_proxy_agent.py: remove 'traditional vs proxy' framing MCPProxyAgent is the standard approach for remote agent integration - Delete workflow_typed_contracts.py: not relevant to MCP integration - README: 7 scripts, renumber hybrid 6-7, remove typed-contract sections * Add LangGraph + MAF GroupChat cross-framework demo (Scripts 8-9) - mcp_server_langgraph.py: LangGraph ReAct agent exposed as MCP server on port 8003 with architecture tools (pattern eval, migration estimate, tech stack recommendation) - workflow_group_chat.py: MAF GroupChatBuilder orchestration with LLM planner routing between local BusinessStrategist (MAF) and remote TechnicalArchitect (LangGraph via MCP) - pyproject.toml: add langgraph, langchain-openai, langchain-core deps - README.md: update to 9 scripts, add cross-framework architecture diagram, Quick Start sections, and dependency table entries * MCP agent demo: rewrite README as MCP-vs-A2A thesis, delete proxy agent, finalize group chat with inline Planner * README: Mermaid diagrams, professional tone, remove LinkedIn teasing language * README: reframe around multi-framework interop problem, two design patterns (Agent-as-Tool + Agent Adapter), add conceptual architecture * README: fix Mermaid diagrams - use br tags instead of \n for line breaks * README: fix A2A comparison table - elicitation supported, structured schemas not * Upgrade agent-framework to 1.0.0rc1 and fastmcp to 3.0.2 Breaking changes migrated: - agent-framework 1.0.0rc1: ChatAgent->Agent, AgentThread->AgentSession, run_stream->run(stream=True), model->default_options(ChatOptions), WorkflowOutputEvent/AgentRunEvent/RequestInfoEvent->unified WorkflowEvent, MagenticBuilder now uses constructor kwargs, orchestrations moved to agent_framework_orchestrations package - fastmcp 3.0.2: removed mcp version pin (now pulled by fastmcp), no code changes needed (fully compatible) Files updated: - agentic_ai/agents/agent_framework/single_agent.py - agentic_ai/agents/agent_framework/multi_agent/handoff_multi_domain_agent.py - agentic_ai/agents/agent_framework/multi_agent/reflection_agent.py - agentic_ai/agents/agent_framework/multi_agent/magentic_group.py - agentic_ai/applications/pyproject.toml (agent-framework==1.0.0rc1) - agentic_ai/workflow/fraud_detection_durable/pyproject.toml (agent-framework==1.0.0rc1) - mcp/pyproject.toml (fastmcp==3.0.2) - tests/test_agent_framework_rc1_regression.py (51 regression tests, all passing) * Fix CI: regenerate requirements.txt with agent-framework-core==1.0.0rc1 The Docker build uses requirements.txt (not pyproject.toml/uv.lock), and it still had agent-framework-core==1.0.0b260130 pinned, which conflicts with agent-framework==1.0.0rc1 requiring agent-framework-core==1.0.0rc1. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22315210980 * Fix: use GH_PAT for auto-merge to trigger downstream workflows Pushes made by GITHUB_TOKEN don't trigger other workflows (GitHub Actions limitation to prevent infinite loops). This meant the auto-merge into int-agentic never triggered promote-to-main.yml. Now uses secrets.GH_PAT with fallback to GITHUB_TOKEN so: 1. If GH_PAT is configured: merge triggers promote-to-main.yml 2. If not: merge still works, but promotion PR must be created manually * Add workflow_dispatch trigger to promote-to-main Allows manual triggering when auto-trigger is missed (e.g. after GITHUB_TOKEN-based merges that don't fire downstream workflows). * Fix: Cosmos DB auth + remove phantom reflection_workflow_agent 1. MCP Cosmos DB: Replace AzureCliCredential (not available in containers) with ManagedIdentityCredential (when AZURE_CLIENT_ID is set) or DefaultAzureCredential as fallback. This fixes the 'Azure CLI not found on path' error in production Container Apps. 2. Backend: Remove reflection_workflow_agent from DEFAULT_AVAILABLE_AGENTS and descriptions - the module file doesn't exist, causing it to show as a broken option in the UI agent selector. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22321497108 * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Trigger rebuild: deploy Cosmos DB auth fix and agent list cleanup * Fix: iterate ResponseStream directly, not .updates ResponseStream is AsyncIterable - iterate with 'async for chunk in stream:' The .updates property returns a Sequence (list), not an async iterator. This caused TypeError: 'async for' requires __aiter__, got list. Not caught in local tests because unit tests mocked the streaming and the regression tests verified API signatures but didn't run live agents. * Fix: make integration test failures block the pipeline Removed continue-on-error: true from the pytest step so test failures actually fail the workflow. Previously, 5/7 tests could fail and the pipeline would still report success and proceed to auto-merge. * Fix: use SHA-tagged images for container deployment, not :latest Container Apps don't create a new revision when the image tag is unchanged. Using :latest meant the old container kept running even after a new image was pushed. Now passes github.sha as the image tag, which forces a new revision on every deploy. This was the root cause of the ResponseStream.updates bug reaching production - the integration tests ran against old containers. * Skip MCP localhost check in CI mode In CI, the evaluation talks to the backend via HTTP which connects to the deployed MCP container internally. The localhost:8000 check is only useful for local dev and was producing a confusing warning in CI logs. * Fix: auto-merge creates PR if none exists Previously auto-merge only looked for an existing open PR and skipped if none was found. Now it auto-creates a PR from the dev branch to int-agentic if one doesn't exist, then merges it. This ensures every successful pipeline run promotes code to int-agentic regardless of whether a PR was manually created beforehand. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22330447642 * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22334055353 * Migrate workshop to agent-framework==1.2.1 (incl. native HandoffBuilder) (#419) * Promote: int-agentic → main (production) (#414) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * Updating Durable Agent Implementation (#404) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * chore: reduce anomaly probability to 1% for controlled demo pace * Reduce anomaly probability to 1% for controlled demo pace (#406) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * chore: reduce anomaly probability to 1% for controlled demo pace --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * add mcp_agent_demo * Add hybrid MCP server (strict-schema + natural-language tools) and typed-contract workflow Scripts 6-8 for the MCP agent demo: - workflow_typed_contracts.py: 4-agent IT security incident response pipeline with Pydantic-enforced contracts at every boundary (SecurityAlert, ThreatAssessment, ImpactAnalysis, IncidentResponse) - mcp_server_hybrid.py: Single MCP endpoint exposing BOTH tool types: * Strict-schema: triage_alert, assess_threat, create_response (Pydantic) * Natural-language: ask_security_advisor, explain_for_customer (prose) * Shared session state across all tools - mcp_client_hybrid.py: 5-step incident flow using both tool types with explicit context passing between steps - README.md: Updated from 6 to 8 capabilities with Hybrid Mode architecture diagram and Quick Start sections 7-8 * Replace workflow_local_remote.py with simplified proxy agent (Script 3) - Delete workflow_local_remote.py (old Script 3 with extra LLM call) - Add workflow_proxy_agent.py: MCPProxyAgent (BaseAgent, no LLM) calls MCP tools directly via call_tool() zero local LLM overhead - Simplify MCPProxyAgent: extract common _call() method, module-level _last_text() helper, remove verbose static methods - Fix workflow_typed_contracts.py: add async with context managers - Update README: merge Script 9 into Script 3 position, update architecture diagram, Quick Start, How It Works sections * Clean up MCP demos: remove comparison commentary, drop typed-contracts - workflow_proxy_agent.py: remove 'traditional vs proxy' framing MCPProxyAgent is the standard approach for remote agent integration - Delete workflow_typed_contracts.py: not relevant to MCP integration - README: 7 scripts, renumber hybrid 6-7, remove typed-contract sections * Add LangGraph + MAF GroupChat cross-framework demo (Scripts 8-9) - mcp_server_langgraph.py: LangGraph ReAct agent exposed as MCP server on port 8003 with architecture tools (pattern eval, migration estimate, tech stack recommendation) - workflow_group_chat.py: MAF GroupChatBuilder orchestration with LLM planner routing between local BusinessStrategist (MAF) and remote TechnicalArchitect (LangGraph via MCP) - pyproject.toml: add langgraph, langchain-openai, langchain-core deps - README.md: update to 9 scripts, add cross-framework architecture diagram, Quick Start sections, and dependency table entries * MCP agent demo: rewrite README as MCP-vs-A2A thesis, delete proxy agent, finalize group chat with inline Planner * README: Mermaid diagrams, professional tone, remove LinkedIn teasing language * README: reframe around multi-framework interop problem, two design patterns (Agent-as-Tool + Agent Adapter), add conceptual architecture * README: fix Mermaid diagrams - use br tags instead of \n for line breaks * README: fix A2A comparison table - elicitation supported, structured schemas not * Upgrade agent-framework to 1.0.0rc1 and fastmcp to 3.0.2 Breaking changes migrated: - agent-framework 1.0.0rc1: ChatAgent->Agent, AgentThread->AgentSession, run_stream->run(stream=True), model->default_options(ChatOptions), WorkflowOutputEvent/AgentRunEvent/RequestInfoEvent->unified WorkflowEvent, MagenticBuilder now uses constructor kwargs, orchestrations moved to agent_framework_orchestrations package - fastmcp 3.0.2: removed mcp version pin (now pulled by fastmcp), no code changes needed (fully compatible) Files updated: - agentic_ai/agents/agent_framework/single_agent.py - agentic_ai/agents/agent_framework/multi_agent/handoff_multi_domain_agent.py - agentic_ai/agents/agent_framework/multi_agent/reflection_agent.py - agentic_ai/agents/agent_framework/multi_agent/magentic_group.py - agentic_ai/applications/pyproject.toml (agent-framework==1.0.0rc1) - agentic_ai/workflow/fraud_detection_durable/pyproject.toml (agent-framework==1.0.0rc1) - mcp/pyproject.toml (fastmcp==3.0.2) - tests/test_agent_framework_rc1_regression.py (51 regression tests, all passing) * Fix CI: regenerate requirements.txt with agent-framework-core==1.0.0rc1 The Docker build uses requirements.txt (not pyproject.toml/uv.lock), and it still had agent-framework-core==1.0.0b260130 pinned, which conflicts with agent-framework==1.0.0rc1 requiring agent-framework-core==1.0.0rc1. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22315210980 * Fix: use GH_PAT for auto-merge to trigger downstream workflows Pushes made by GITHUB_TOKEN don't trigger other workflows (GitHub Actions limitation to prevent infinite loops). This meant the auto-merge into int-agentic never triggered promote-to-main.yml. Now uses secrets.GH_PAT with fallback to GITHUB_TOKEN so: 1. If GH_PAT is configured: merge triggers promote-to-main.yml 2. If not: merge still works, but promotion PR must be created manually * Add workflow_dispatch trigger to promote-to-main Allows manual triggering when auto-trigger is missed (e.g. after GITHUB_TOKEN-based merges that don't fire downstream workflows). * Fix: Cosmos DB auth + remove phantom reflection_workflow_agent 1. MCP Cosmos DB: Replace AzureCliCredential (not available in containers) with ManagedIdentityCredential (when AZURE_CLIENT_ID is set) or DefaultAzureCredential as fallback. This fixes the 'Azure CLI not found on path' error in production Container Apps. 2. Backend: Remove reflection_workflow_agent from DEFAULT_AVAILABLE_AGENTS and descriptions - the module file doesn't exist, causing it to show as a broken option in the UI agent selector. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22321497108 * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production …

…2.1 storages (#423) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * Updating Durable Agent Implementation (#404) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * chore: reduce anomaly probability to 1% for controlled demo pace * Reduce anomaly probability to 1% for controlled demo pace (#406) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * chore: reduce anomaly probability to 1% for controlled demo pace --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * add mcp_agent_demo * Add hybrid MCP server (strict-schema + natural-language tools) and typed-contract workflow Scripts 6-8 for the MCP agent demo: - workflow_typed_contracts.py: 4-agent IT security incident response pipeline with Pydantic-enforced contracts at every boundary (SecurityAlert, ThreatAssessment, ImpactAnalysis, IncidentResponse) - mcp_server_hybrid.py: Single MCP endpoint exposing BOTH tool types: * Strict-schema: triage_alert, assess_threat, create_response (Pydantic) * Natural-language: ask_security_advisor, explain_for_customer (prose) * Shared session state across all tools - mcp_client_hybrid.py: 5-step incident flow using both tool types with explicit context passing between steps - README.md: Updated from 6 to 8 capabilities with Hybrid Mode architecture diagram and Quick Start sections 7-8 * Replace workflow_local_remote.py with simplified proxy agent (Script 3) - Delete workflow_local_remote.py (old Script 3 with extra LLM call) - Add workflow_proxy_agent.py: MCPProxyAgent (BaseAgent, no LLM) calls MCP tools directly via call_tool() zero local LLM overhead - Simplify MCPProxyAgent: extract common _call() method, module-level _last_text() helper, remove verbose static methods - Fix workflow_typed_contracts.py: add async with context managers - Update README: merge Script 9 into Script 3 position, update architecture diagram, Quick Start, How It Works sections * Clean up MCP demos: remove comparison commentary, drop typed-contracts - workflow_proxy_agent.py: remove 'traditional vs proxy' framing MCPProxyAgent is the standard approach for remote agent integration - Delete workflow_typed_contracts.py: not relevant to MCP integration - README: 7 scripts, renumber hybrid 6-7, remove typed-contract sections * Add LangGraph + MAF GroupChat cross-framework demo (Scripts 8-9) - mcp_server_langgraph.py: LangGraph ReAct agent exposed as MCP server on port 8003 with architecture tools (pattern eval, migration estimate, tech stack recommendation) - workflow_group_chat.py: MAF GroupChatBuilder orchestration with LLM planner routing between local BusinessStrategist (MAF) and remote TechnicalArchitect (LangGraph via MCP) - pyproject.toml: add langgraph, langchain-openai, langchain-core deps - README.md: update to 9 scripts, add cross-framework architecture diagram, Quick Start sections, and dependency table entries * MCP agent demo: rewrite README as MCP-vs-A2A thesis, delete proxy agent, finalize group chat with inline Planner * README: Mermaid diagrams, professional tone, remove LinkedIn teasing language * README: reframe around multi-framework interop problem, two design patterns (Agent-as-Tool + Agent Adapter), add conceptual architecture * README: fix Mermaid diagrams - use br tags instead of \n for line breaks * README: fix A2A comparison table - elicitation supported, structured schemas not * Upgrade agent-framework to 1.0.0rc1 and fastmcp to 3.0.2 Breaking changes migrated: - agent-framework 1.0.0rc1: ChatAgent->Agent, AgentThread->AgentSession, run_stream->run(stream=True), model->default_options(ChatOptions), WorkflowOutputEvent/AgentRunEvent/RequestInfoEvent->unified WorkflowEvent, MagenticBuilder now uses constructor kwargs, orchestrations moved to agent_framework_orchestrations package - fastmcp 3.0.2: removed mcp version pin (now pulled by fastmcp), no code changes needed (fully compatible) Files updated: - agentic_ai/agents/agent_framework/single_agent.py - agentic_ai/agents/agent_framework/multi_agent/handoff_multi_domain_agent.py - agentic_ai/agents/agent_framework/multi_agent/reflection_agent.py - agentic_ai/agents/agent_framework/multi_agent/magentic_group.py - agentic_ai/applications/pyproject.toml (agent-framework==1.0.0rc1) - agentic_ai/workflow/fraud_detection_durable/pyproject.toml (agent-framework==1.0.0rc1) - mcp/pyproject.toml (fastmcp==3.0.2) - tests/test_agent_framework_rc1_regression.py (51 regression tests, all passing) * Fix CI: regenerate requirements.txt with agent-framework-core==1.0.0rc1 The Docker build uses requirements.txt (not pyproject.toml/uv.lock), and it still had agent-framework-core==1.0.0b260130 pinned, which conflicts with agent-framework==1.0.0rc1 requiring agent-framework-core==1.0.0rc1. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22315210980 * Fix: use GH_PAT for auto-merge to trigger downstream workflows Pushes made by GITHUB_TOKEN don't trigger other workflows (GitHub Actions limitation to prevent infinite loops). This meant the auto-merge into int-agentic never triggered promote-to-main.yml. Now uses secrets.GH_PAT with fallback to GITHUB_TOKEN so: 1. If GH_PAT is configured: merge triggers promote-to-main.yml 2. If not: merge still works, but promotion PR must be created manually * Add workflow_dispatch trigger to promote-to-main Allows manual triggering when auto-trigger is missed (e.g. after GITHUB_TOKEN-based merges that don't fire downstream workflows). * Fix: Cosmos DB auth + remove phantom reflection_workflow_agent 1. MCP Cosmos DB: Replace AzureCliCredential (not available in containers) with ManagedIdentityCredential (when AZURE_CLIENT_ID is set) or DefaultAzureCredential as fallback. This fixes the 'Azure CLI not found on path' error in production Container Apps. 2. Backend: Remove reflection_workflow_agent from DEFAULT_AVAILABLE_AGENTS and descriptions - the module file doesn't exist, causing it to show as a broken option in the UI agent selector. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22321497108 * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Trigger rebuild: deploy Cosmos DB auth fix and agent list cleanup * Fix: iterate ResponseStream directly, not .updates ResponseStream is AsyncIterable - iterate with 'async for chunk in stream:' The .updates property returns a Sequence (list), not an async iterator. This caused TypeError: 'async for' requires __aiter__, got list. Not caught in local tests because unit tests mocked the streaming and the regression tests verified API signatures but didn't run live agents. * Fix: make integration test failures block the pipeline Removed continue-on-error: true from the pytest step so test failures actually fail the workflow. Previously, 5/7 tests could fail and the pipeline would still report success and proceed to auto-merge. * Fix: use SHA-tagged images for container deployment, not :latest Container Apps don't create a new revision when the image tag is unchanged. Using :latest meant the old container kept running even after a new image was pushed. Now passes github.sha as the image tag, which forces a new revision on every deploy. This was the root cause of the ResponseStream.updates bug reaching production - the integration tests ran against old containers. * Skip MCP localhost check in CI mode In CI, the evaluation talks to the backend via HTTP which connects to the deployed MCP container internally. The localhost:8000 check is only useful for local dev and was producing a confusing warning in CI logs. * Fix: auto-merge creates PR if none exists Previously auto-merge only looked for an existing open PR and skipped if none was found. Now it auto-creates a PR from the dev branch to int-agentic if one doesn't exist, then merges it. This ensures every successful pipeline run promotes code to int-agentic regardless of whether a PR was manually created beforehand. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22330447642 * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22334055353 * Migrate workshop to agent-framework==1.2.1 (incl. native HandoffBuilder) (#419) * Promote: int-agentic → main (production) (#414) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * Updating Durable Agent Implementation (#404) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * chore: reduce anomaly probability to 1% for controlled demo pace * Reduce anomaly probability to 1% for controlled demo pace (#406) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * chore: reduce anomaly probability to 1% for controlled demo pace --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * add mcp_agent_demo * Add hybrid MCP server (strict-schema + natural-language tools) and typed-contract workflow Scripts 6-8 for the MCP agent demo: - workflow_typed_contracts.py: 4-agent IT security incident response pipeline with Pydantic-enforced contracts at every boundary (SecurityAlert, ThreatAssessment, ImpactAnalysis, IncidentResponse) - mcp_server_hybrid.py: Single MCP endpoint exposing BOTH tool types: * Strict-schema: triage_alert, assess_threat, create_response (Pydantic) * Natural-language: ask_security_advisor, explain_for_customer (prose) * Shared session state across all tools - mcp_client_hybrid.py: 5-step incident flow using both tool types with explicit context passing between steps - README.md: Updated from 6 to 8 capabilities with Hybrid Mode architecture diagram and Quick Start sections 7-8 * Replace workflow_local_remote.py with simplified proxy agent (Script 3) - Delete workflow_local_remote.py (old Script 3 with extra LLM call) - Add workflow_proxy_agent.py: MCPProxyAgent (BaseAgent, no LLM) calls MCP tools directly via call_tool() zero local LLM overhead - Simplify MCPProxyAgent: extract common _call() method, module-level _last_text() helper, remove verbose static methods - Fix workflow_typed_contracts.py: add async with context managers - Update README: merge Script 9 into Script 3 position, update architecture diagram, Quick Start, How It Works sections * Clean up MCP demos: remove comparison commentary, drop typed-contracts - workflow_proxy_agent.py: remove 'traditional vs proxy' framing MCPProxyAgent is the standard approach for remote agent integration - Delete workflow_typed_contracts.py: not relevant to MCP integration - README: 7 scripts, renumber hybrid 6-7, remove typed-contract sections * Add LangGraph + MAF GroupChat cross-framework demo (Scripts 8-9) - mcp_server_langgraph.py: LangGraph ReAct agent exposed as MCP server on port 8003 with architecture tools (pattern eval, migration estimate, tech stack recommendation) - workflow_group_chat.py: MAF GroupChatBuilder orchestration with LLM planner routing between local BusinessStrategist (MAF) and remote TechnicalArchitect (LangGraph via MCP) - pyproject.toml: add langgraph, langchain-openai, langchain-core deps - README.md: update to 9 scripts, add cross-framework architecture diagram, Quick Start sections, and dependency table entries * MCP agent demo: rewrite README as MCP-vs-A2A thesis, delete proxy agent, finalize group chat with inline Planner * README: Mermaid diagrams, professional tone, remove LinkedIn teasing language * README: reframe around multi-framework interop problem, two design patterns (Agent-as-Tool + Agent Adapter), add conceptual architecture * README: fix Mermaid diagrams - use br tags instead of \n for line breaks * README: fix A2A comparison table - elicitation supported, structured schemas not * Upgrade agent-framework to 1.0.0rc1 and fastmcp to 3.0.2 Breaking changes migrated: - agent-framework 1.0.0rc1: ChatAgent->Agent, AgentThread->AgentSession, run_stream->run(stream=True), model->default_options(ChatOptions), WorkflowOutputEvent/AgentRunEvent/RequestInfoEvent->unified WorkflowEvent, MagenticBuilder now uses constructor kwargs, orchestrations moved to agent_framework_orchestrations package - fastmcp 3.0.2: removed mcp version pin (now pulled by fastmcp), no code changes needed (fully compatible) Files updated: - agentic_ai/agents/agent_framework/single_agent.py - agentic_ai/agents/agent_framework/multi_agent/handoff_multi_domain_agent.py - agentic_ai/agents/agent_framework/multi_agent/reflection_agent.py - agentic_ai/agents/agent_framework/multi_agent/magentic_group.py - agentic_ai/applications/pyproject.toml (agent-framework==1.0.0rc1) - agentic_ai/workflow/fraud_detection_durable/pyproject.toml (agent-framework==1.0.0rc1) - mcp/pyproject.toml (fastmcp==3.0.2) - tests/test_agent_framework_rc1_regression.py (51 regression tests, all passing) * Fix CI: regenerate requirements.txt with agent-framework-core==1.0.0rc1 The Docker build uses requirements.txt (not pyproject.toml/uv.lock), and it still had agent-framework-core==1.0.0b260130 pinned, which conflicts with agent-framework==1.0.0rc1 requiring agent-framework-core==1.0.0rc1. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22315210980 * Fix: use GH_PAT for auto-merge to trigger downstream workflows Pushes made by GITHUB_TOKEN don't trigger other workflows (GitHub Actions limitation to prevent infinite loops). This meant the auto-merge into int-agentic never triggered promote-to-main.yml. Now uses secrets.GH_PAT with fallback to GITHUB_TOKEN so: 1. If GH_PAT is configured: merge triggers promote-to-main.yml 2. If not: merge still works, but promotion PR must be created manually * Add workflow_dispatch trigger to promote-to-main Allows manual triggering when auto-trigger is missed (e.g. after GITHUB_TOKEN-based merges that don't fire downstream workflows). * Fix: Cosmos DB auth + remove phantom reflection_workflow_agent 1. MCP Cosmos DB: Replace AzureCliCredential (not available in containers) with ManagedIdentityCredential (when AZURE_CLIENT_ID is set) or DefaultAzureCredential as fallback. This fixes the 'Azure CLI not found on path' error in production Container Apps. 2. Backend: Remove reflection_workflow_agent from DEFAULT_AVAILABLE_AGENTS and descriptions - the module file doesn't exist, causing it to show as a broken option in the UI agent selector. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22321497108 * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Trigger rebuild: deploy Cosmos DB auth fix and agent list cleanup * Fix: iterate ResponseStream directly, not .updates ResponseStream is AsyncIterable - iterate with 'async for chunk in stream:' The .updates property returns a Sequence (list), not an async iterator. This caused TypeError: 'async for' requires __aiter__, got list. Not caught in local tests because unit tests mocked the streaming and the regression tests verified API signatures but didn't run live agents. * Fix: make integration test failures block the pipeline Removed continue-on-error: true from the pytest step so test failures actually fail the workflow. Previously, 5/7 tests could fail and the pipeline would still report success and proceed to auto-merge. * Fix: use SHA-tagged images for container deployment, not :latest Container Apps don't create a new revision when the image tag is unchanged. Using :latest meant the old container kept running even after a new image was pushed. Now passes github.sha as the image tag, which forces a new revision on every deploy. This was the root cause of the ResponseStream.updates bug reaching production - the integration tests ran against old containers. * Skip MCP localhost check in CI mode In CI, the evaluation talks to the backend via HTTP which connects to the deployed MCP container internally. The localhost:8000 check is only useful for local dev and was producing a confusing warning in CI logs. * Fix: auto-merge creates PR if none exists Previously auto-merge only looked for an existing open PR and skipped if none was found. Now it auto-creates a PR from the dev branch to int-agentic if one doesn't exist, then merges it. This ensures every successful pipeline run promotes code to int-agentic regardless of whether a PR was manually created beforehand. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22330447642 --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * Promote: int-agentic → main (production) (#416) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * Updating Durable Agent Implementation (#404) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * chore: reduce anomaly probability to 1% for controlled demo pace * Reduce anomaly probability to 1% for controlled demo pace (#406) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * chore: reduce anomaly probability to 1% for controlled demo pace --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * add mcp_agent_demo * Add hybrid MCP server (strict-schema + natural-language tools) and typed-contract workflow Scripts 6-8 for the MCP agent demo: - workflow_typed_contracts.py: 4-agent IT security incident response pipeline with Pydantic-enforced contracts at every boundary (SecurityAlert, ThreatAssessment, ImpactAnalysis, IncidentResponse) - mcp_server_hybrid.py: Single MCP endpoint exposing BOTH tool types: * Strict-schema: triage_alert, assess_threat, create_response (Pydantic) * Natural-language: ask_security_advisor, explain_for_customer (prose) * Shared session state across all tools - mcp_client_hybrid.py: 5-step incident flow using both tool types with explicit context passing between steps - README.md: Updated from 6 to 8 capabilities with Hybrid Mode architecture diagram and Quick Start sections 7-8 * Replace workflow_local_remote.py with simplified proxy agent (Script 3) - Delete workflow_local_remote.py (old Script 3 with extra LLM call) - Add workflow_proxy_agent.py: MCPProxyAgent (BaseAgent, no LLM) calls MCP tools directly via call_tool() zero local LLM overhead - Simplify MCPProxyAgent: extract common _call() method, module-level _last_text() helper, remove verbose static methods - Fix workflow_typed_contracts.py: add async with context managers - Update README: merge Script 9 into Script 3 position, update architecture diagram, Quick Start, How It Works sections * Clean up MCP demos: remove comparison commentary, drop typed-contracts - workflow_proxy_agent.py: remove 'traditional vs proxy' framing MCPProxyAgent is the standard approach for remote agent integration - Delete workflow_typed_contracts.py: not relevant to MCP integration - README: 7 scripts, renumber hybrid 6-7, remove typed-contract sections * Add LangGraph + MAF GroupChat cross-framework demo (Scripts 8-9) - mcp_server_langgraph.py: LangGraph ReAct agent exposed as MCP server on port 8003 with architecture tools (pattern eval, migration estimate, tech stack recommendation) - workflow_group_chat.py: MAF GroupChatBuilder orchestration with LLM planner routing between local BusinessStrategist (MAF) and remote TechnicalArchitect (LangGraph via MCP) - pyproject.toml: add langgraph, langchain-openai, langchain-core deps - README.md: update to 9 scripts, add cross-framework architecture diagram, Quick Start sections, and dependency table entries * MCP agent demo: rewrite README as MCP-vs-A2A thesis, delete proxy agent, finalize group chat with inline Planner * README: Mermaid diagrams, professional tone, remove LinkedIn teasing language * README: reframe around multi-framework interop problem, two design patterns (Agent-as-Tool + Agent Adapter), add conceptual architecture * README: fix Mermaid diagrams - use br tags instead of \n for line breaks * README: fix A2A comparison table - elicitation supported, structured schemas not * Upgrade agent-framework to 1.0.0rc1 and fastmcp to 3.0.2 Breaking changes migrated: - agent-framework 1.0.0rc1: ChatAgent->Agent, AgentThread->AgentSession, run_stream->run(stream=True), model->default_options(ChatOptions), WorkflowOutputEvent/AgentRunEvent/RequestInfoEvent->unified WorkflowEvent, MagenticBuilder now uses constructor kwargs, orchestrations moved to agent_framework_orchestrations package - fastmcp 3.0.2: removed mcp version pin (now pulled by fastmcp), no code changes needed (fully compatible) Files updated: - agentic_ai/agents/agent_framework/single_agent.py - agentic_ai/agents/agent_framework/multi_agent/handoff_multi_domain_agent.py - agentic_ai/agents/agent_framework/multi_agent/reflection_agent.py - agentic_ai/agents/agent_framework/multi_agent/magentic_group.py - agentic_ai/applications/pyproject.toml (agent-framework==1.0.0rc1) - agentic_ai/workflow/fraud_detection_durable/pyproject.toml (agent-framework==1.0.0rc1) - mcp/pyproject.toml (fastmcp==3.0.2) - tests/test_agent_framework_rc1_regression.py (51 regression tests, all passing) * Fix CI: regenerate requirements.txt with agent-framework-core==1.0.0rc1 The Docker build uses requirements.txt (not pyproject.toml/uv.lock), and it still had agent-framework-core==1.0.0b260130 pinned, which conflicts with agent-framework==1.0.0rc1 requiring agent-framework-core==1.0.0rc1. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22315210980 * Fix: use GH_PAT for auto-merge to trigger downstream workflows Pushes made by GITHUB_TOKEN don't trigger other workflows (GitHub Actions limitation to prevent infinite loops). This meant the auto-merge into int-agentic never triggered promote-to-main.yml. Now uses secrets.GH_PAT with fallback to GITHUB_TOKEN so: 1. If GH_PAT is configured: merge triggers promote-to-main.yml 2. If not: merge still works, but promotion PR must be created manually * Add workflow_dispatch trigger to promote-to-main Allows manual triggering when auto-trigger is missed (e.g. after GITHUB_TOKEN-based merges that don't fire downstream workflows). * Fix: Cosmos DB auth + remove phantom reflection_workflow_agent 1. MCP Cosmos DB: Replace AzureCliCredential (not available in containers) with ManagedIdentityCredential (when AZURE_CLIENT_ID is set) or DefaultAzureCredential as fallback. This fixes the 'Azure CLI not found on path' error in production Container Apps. 2. Backend: Remove reflection_workflow_agent from DEFAULT_AVAILABLE_AGENTS and descriptions - the module file doesn't exist, causing it to show as a broken option in the UI agent selector. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22321497108 * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Trigger rebuild: deploy Cosmos DB auth fix and agent list cleanup * Fix: iterate ResponseStream directly, not .updates ResponseStream is AsyncIterable - iterate with 'async for chunk in stream:' The .updates property returns a Sequence (list), not an async iterator. This caused TypeError: 'async for' requires __aiter__, got list. Not caught in local tests because unit tests mocked the streaming and the regression tests verified API signatures but didn't run live agents. * Fix: make integration test failures block the pipeline Removed continue-on-error: true from the pytest step so test failures actually fail the workflow. Previously, 5/7 tests could fail and the pipeline would still report success and proceed to auto-merge. * Fix: use SHA-tagged images for container deployment, not :latest Container Apps don't create a new revision when the image tag is unchanged. Using :latest meant the old container kept running even after a new image was pushed. Now passes github.sha as the image tag, which forces a new revision on every deploy. This was the root cause of the ResponseStream.updates bug reaching production - the integration tests ran against old containers. * Skip MCP localhost check in CI mode In CI, the evaluation talks to the backend via HTTP which connects to the deployed MCP container internally. The localhost:8000 check is only useful for local dev and was producing a confusing warning in CI logs. * Fix: auto-merge creates PR if none exists Previously auto-merge only looked for an existing open PR and skipped if none was found. Now it auto-creates a PR from the dev branch to int-agentic if one doesn't exist, then merges it. This ensures every successful pipeline run promotes code to int-agentic regardless of whether a PR was manually created beforehand. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22330447642 * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22334055353 --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * Bump agent-framework to 1.2.1 and migrate AzureOpenAIChatClient → OpenAIChatClient Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> * Rewrite handoff agent on native HandoffBuilder; update regression tests for 1.2.1 Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> * Update HANDOFF_README and dependency table for native HandoffBuilder migration Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> * Address GPT-5.5 review: fix CheckpointStorage protocol, reflection_agent kwargs, observability sample Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> * Fix PEP 8 spacing in mcp_agent_demo files (code review nits) Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> * Regenerate uv.lock files; pass workflow_name to 1.2.x checkpoint listing Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/91307ce7-2444-4b5f-b75f-2b736bc022c5 Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> --------- Co-authored-by: James Nguyen <janguy@microsoft.com> Co-authored-by: James N. <james.nguyen@microsoft.com> Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com> Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> * docs: Remove misleading RequestInfoExecutor references (#418) doc update * Replace custom DictCheckpointStorage with built-in 1.2.1 storages Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/f6da6e3e-b288-425d-ac4d-c7ba6fe62fe7 Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> * tests: use asyncio.run() instead of new/get_event_loop() Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/f6da6e3e-b288-425d-ac4d-c7ba6fe62fe7 Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com> * Fix backend 500s by renaming ChatOptions(model_id=…) to ChatOptions(model=…) for agent-framework 1.2.1 (#421) * Promote: int-agentic → main (production) (#414) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * Updating Durable Agent Implementation (#404) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * chore: reduce anomaly probability to 1% for controlled demo pace * Reduce anomaly probability to 1% for controlled demo pace (#406) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * chore: reduce anomaly probability to 1% for controlled demo pace --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * add mcp_agent_demo * Add hybrid MCP server (strict-schema + natural-language tools) and typed-contract workflow Scripts 6-8 for the MCP agent demo: - workflow_typed_contracts.py: 4-agent IT security incident response pipeline with Pydantic-enforced contracts at every boundary (SecurityAlert, ThreatAssessment, ImpactAnalysis, IncidentResponse) - mcp_server_hybrid.py: Single MCP endpoint exposing BOTH tool types: * Strict-schema: triage_alert, assess_threat, create_response (Pydantic) * Natural-language: ask_security_advisor, explain_for_customer (prose) * Shared session state across all tools - mcp_client_hybrid.py: 5-step incident flow using both tool types with explicit context passing between steps - README.md: Updated from 6 to 8 capabilities with Hybrid Mode architecture diagram and Quick Start sections 7-8 * Replace workflow_local_remote.py with simplified proxy agent (Script 3) - Delete workflow_local_remote.py (old Script 3 with extra LLM call) - Add workflow_proxy_agent.py: MCPProxyAgent (BaseAgent, no LLM) calls MCP tools directly via call_tool() zero local LLM overhead - Simplify MCPProxyAgent: extract common _call() method, module-level _last_text() helper, remove verbose static methods - Fix workflow_typed_contracts.py: add async with context managers - Update README: merge Script 9 into Script 3 position, update architecture diagram, Quick Start, How It Works sections * Clean up MCP demos: remove comparison commentary, drop typed-contracts - workflow_proxy_agent.py: remove 'traditional vs proxy' framing MCPProxyAgent is the standard approach for remote agent integration - Delete workflow_typed_contracts.py: not relevant to MCP integration - README: 7 scripts, renumber hybrid 6-7, remove typed-contract sections * Add LangGraph + MAF GroupChat cross-framework demo (Scripts 8-9) - mcp_server_langgraph.py: LangGraph ReAct agent exposed as MCP server on port 8003 with architecture tools (pattern eval, migration estimate, tech stack recommendation) - workflow_group_chat.py: MAF GroupChatBuilder orchestration with LLM planner routing between local BusinessStrategist (MAF) and remote TechnicalArchitect (LangGraph via MCP) - pyproject.toml: add langgraph, langchain-openai, langchain-core deps - README.md: update to 9 scripts, add cross-framework architecture diagram, Quick Start sections, and dependency table entries * MCP agent demo: rewrite README as MCP-vs-A2A thesis, delete proxy agent, finalize group chat with inline Planner * README: Mermaid diagrams, professional tone, remove LinkedIn teasing language * README: reframe around multi-framework interop problem, two design patterns (Agent-as-Tool + Agent Adapter), add conceptual architecture * README: fix Mermaid diagrams - use br tags instead of \n for line breaks * README: fix A2A comparison table - elicitation supported, structured schemas not * Upgrade agent-framework to 1.0.0rc1 and fastmcp to 3.0.2 Breaking changes migrated: - agent-framework 1.0.0rc1: ChatAgent->Agent, AgentThread->AgentSession, run_stream->run(stream=True), model->default_options(ChatOptions), WorkflowOutputEvent/AgentRunEvent/RequestInfoEvent->unified WorkflowEvent, MagenticBuilder now uses constructor kwargs, orchestrations moved to agent_framework_orchestrations package - fastmcp 3.0.2: removed mcp version pin (now pulled by fastmcp), no code changes needed (fully compatible) Files updated: - agentic_ai/agents/agent_framework/single_agent.py - agentic_ai/agents/agent_framework/multi_agent/handoff_multi_domain_agent.py - agentic_ai/agents/agent_framework/multi_agent/reflection_agent.py - agentic_ai/agents/agent_framework/multi_agent/magentic_group.py - agentic_ai/applications/pyproject.toml (agent-framework==1.0.0rc1) - agentic_ai/workflow/fraud_detection_durable/pyproject.toml (agent-framework==1.0.0rc1) - mcp/pyproject.toml (fastmcp==3.0.2) - tests/test_agent_framework_rc1_regression.py (51 regression tests, all passing) * Fix CI: regenerate requirements.txt with agent-framework-core==1.0.0rc1 The Docker build uses requirements.txt (not pyproject.toml/uv.lock), and it still had agent-framework-core==1.0.0b260130 pinned, which conflicts with agent-framework==1.0.0rc1 requiring agent-framework-core==1.0.0rc1. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22315210980 * Fix: use GH_PAT for auto-merge to trigger downstream workflows Pushes made by GITHUB_TOKEN don't trigger other workflows (GitHub Actions limitation to prevent infinite loops). This meant the auto-merge into int-agentic never triggered promote-to-main.yml. Now uses secrets.GH_PAT with fallback to GITHUB_TOKEN so: 1. If GH_PAT is configured: merge triggers promote-to-main.yml 2. If not: merge still works, but promotion PR must be created manually * Add workflow_dispatch trigger to promote-to-main Allows manual triggering when auto-trigger is missed (e.g. after GITHUB_TOKEN-based merges that don't fire downstream workflows). * Fix: Cosmos DB auth + remove phantom reflection_workflow_agent 1. MCP Cosmos DB: Replace AzureCliCredential (not available in containers) with ManagedIdentityCredential (when AZURE_CLIENT_ID is set) or DefaultAzureCredential as fallback. This fixes the 'Azure CLI not found on path' error in production Container Apps. 2. Backend: Remove reflection_workflow_agent from DEFAULT_AVAILABLE_AGENTS and descriptions - the module file doesn't exist, causing it to show as a broken option in the UI agent selector. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22321497108 * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Trigger rebuild: deploy Cosmos DB auth fix and agent list cleanup * Fix: iterate ResponseStream directly, not .updates ResponseStream is AsyncIterable - iterate with 'async for chunk in stream:' The .updates property returns a Sequence (list), not an async iterator. This caused TypeError: 'async for' requires __aiter__, got list. Not caught in local tests because unit tests mocked the streaming and the regression tests verified API signatures but didn't run live agents. * Fix: make integration test failures block the pipeline Removed continue-on-error: true from the pytest step so test failures actually fail the workflow. Previously, 5/7 tests could fail and the pipeline would still report success and proceed to auto-merge. * Fix: use SHA-tagged images for container deployment, not :latest Container Apps don't create a new revision when the image tag is unchanged. Using :latest meant the old container kept running even after a new image was pushed. Now passes github.sha as the image tag, which forces a new revision on every deploy. This was the root cause of the ResponseStream.updates bug reaching production - the integration tests ran against old containers. * Skip MCP localhost check in CI mode In CI, the evaluation talks to the backend via HTTP which connects to the deployed MCP container internally. The localhost:8000 check is only useful for local dev and was producing a confusing warning in CI logs. * Fix: auto-merge creates PR if none exists Previously auto-merge only looked for an existing open PR and skipped if none was found. Now it auto-creates a PR from the dev branch to int-agentic if one doesn't exist, then merges it. This ensures every successful pipeline run promotes code to int-agentic regardless of whether a PR was manually created beforehand. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22330447642 --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * Promote: int-agentic → main (production) (#416) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * Updating Durable Agent Implementation (#404) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * chore: reduce anomaly probability to 1% for controlled demo pace * Reduce anomaly probability to 1% for controlled demo pace (#406) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * chore: reduce anomaly probability to 1% for controlled demo pace --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * add mcp_agent_demo * Add hybrid MCP server (strict-schema + natural-language tools) and typed-contract workflow Scripts 6-8 for the MCP agent demo: - workflow_typed_contracts.py: 4-agent IT security incident response pipeline with Pydantic-enforced contracts at every boundary (SecurityAlert, ThreatAssessment, ImpactAnalysis, IncidentResponse) - mcp_server_hybrid.py: Single MCP endpoint exposing BOTH tool types: * Strict-schema: triage_alert, assess_threat, create_response (Pydantic) * Natural-language: ask_security_advisor, explain_for_customer (prose) * Shared session state across all tools - mcp_client_hybrid.py: 5-step incident flow using both tool types with explicit context passing between steps - README.md: Updated from 6 to 8 capabilities with Hybrid Mode architecture diagram and Quick Start sections 7-8 * Replace workflow_local_remote.py with simplified proxy agent (Script 3) - Delete workflow_local_remote.py (old Script 3 with extra LLM call) - Add workflow_proxy_agent.py: MCPProxyAgent (BaseAgent, no LLM) calls MCP tools directly via call_tool() zero local LLM overhead - Simplify MCPProxyAgent: extract common _call() method, module-level _last_text() helper, remove verbose static methods - Fix workflow_typed_contracts.py: add async with context managers - Update README: merge Script 9 into Script 3 position, update architecture diagram, Quick Start, How It Works sections * Clean up MCP demos: remove comparison commentary, drop typed-contracts - workflow_proxy_agent.py: remove 'traditional vs proxy' framing MCPProxyAgent is the standard approach for remote agent integration - Delete workflow_typed_contracts.py: not relevant to MCP integration - README: 7 scripts, renumber hybrid 6-7, remove typed-contract sections * Add LangGraph + MAF GroupChat cross-framework demo (Scripts 8-9) - mcp_server_langgraph.py: LangGraph ReAct agent exposed as MCP server on port 8003 with architecture tools (pattern eval, migration estimate, tech stack recommendation) - workflow_group_chat.py: MAF GroupChatBuilder orchestration with LLM planner routing between local BusinessStrategist (MAF) and remote TechnicalArchitect (LangGraph via MCP) - pyproject.toml: add langgraph, langchain-openai, langchain-core deps - README.md: update to 9 scripts, add cross-framework architecture diagram, Quick Start sections, and dependency table entries * MCP agent demo: rewrite README as MCP-vs-A2A thesis, delete proxy agent, finalize group chat with inline Planner * README: Mermaid diagrams, professional tone, remove LinkedIn teasing language * README: reframe around multi-framework interop problem, two design patterns (Agent-as-Tool + Agent Adapter), add conceptual architecture * README: fix Mermaid diagrams - use br tags instead of \n for line breaks * README: fix A2A comparison table - elicitation supported, structured schemas not * Upgrade agent-framework to 1.0.0rc1 and fastmcp to 3.0.2 Breaking changes migrated: - agent-framework 1.0.0rc1: ChatAgent->Agent, AgentThread->AgentSession, run_stream->run(stream=True), model->default_options(ChatOptions), WorkflowOutputEvent/AgentRunEvent/RequestInfoEvent->unified WorkflowEvent, MagenticBuilder now uses constructor kwargs, orchestrations moved to agent_framework_orchestrations package - fastmcp 3.0.2: removed mcp version pin (now pulled by fastmcp), no code changes needed (fully compatible) Files updated: - agentic_ai/agents/agent_framework/single_agent.py - agentic_ai/agents/agent_framework/multi_agent/handoff_multi_domain_agent.py - agentic_ai/agents/agent_framework/multi_agent/reflection_agent.py - agentic_ai/agents/agent_framework/multi_agent/magentic_group.py - agentic_ai/applications/pyproject.toml (agent-framework==1.0.0rc1) - agentic_ai/workflow/fraud_detection_durable/pyproject.toml (agent-framework==1.0.0rc1) - mcp/pyproject.toml (fastmcp==3.0.2) - tests/test_agent_framework_rc1_regression.py (51 regression tests, all passing) * Fix CI: regenerate requirements.txt with agent-framework-core==1.0.0rc1 The Docker build uses requirements.txt (not pyproject.toml/uv.lock), and it still had agent-framework-core==1.0.0b260130 pinned, which conflicts with agent-framework==1.0.0rc1 requiring agent-framework-core==1.0.0rc1. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22315210980 * Fix: use GH_PAT for auto-merge to trigger downstream workflows Pushes made by GITHUB_TOKEN don't trigger other workflows (GitHub Actions limitation to prevent infinite loops). This meant the auto-merge into int-agentic never triggered promote-to-main.yml. Now uses secrets.GH_PAT with fallback to GITHUB_TOKEN so: 1. If GH_PAT is configured: merge triggers promote-to-main.yml 2. If not: merge still works, but promotion PR must be created manually * Add workflow_dispatch trigger to promote-to-main Allows manual triggering when auto-trigger is missed (e.g. after GITHUB_TOKEN-based merges that don't fire downstream workflows). * Fix: Cosmos DB auth + remove phantom reflection_workflow_agent 1. MCP Cosmos DB: Replace AzureCliCredential (not available in containers) with ManagedIdentityCredential (when AZURE_CLIENT_ID is set) or DefaultAzureCredential as fallback. This fixes the 'Azure CLI not found on path' error in production Container Apps. 2. Backend: Remove reflection_workflow_agent from DEFAULT_AVAILABLE_AGENTS and descriptions - the module file doesn't exist, causing it to show as a broken option in the UI agent selector. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22321497108 * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Trigger rebuild: deploy Cosmos DB auth fix and agent list cleanup * Fix: iterate ResponseStream directly, not .updates ResponseStream is AsyncIterable - iterate with 'async for chunk in stream:' The .updates property returns a Sequence (list), not an async iterator. This caused TypeError: 'async for' requires __aiter__, got list. Not caught in local tests because unit tests mocked the streaming and the regression tests verified API signatures but didn't run live agents. * Fix: make integration test failures block the pipeline Removed continue-on-error: true from the pytest step so test failures actually fail the workflow. Previously, 5/7 tests could fail and the pipeline would still report success and proceed to auto-merge. * Fix: use SHA-tagged images for container deployment, not :latest Container Apps don't create a new revision when the image tag is unchanged. Using :latest meant the old container kept running even after a new image was pushed. Now passes github.sha as the image tag, which forces a new revision on every deploy. This was the root cause of the ResponseStream.updates bug reaching production - the integration tests ran against old containers. * Skip MCP localhost check in CI mode In CI, the evaluation talks to the backend via HTTP which connects to the deployed MCP container internally. The localhost:8000 check is only useful for local dev and was producing a confusing warning in CI logs. * Fix: auto-merge creates PR if none exists Previously auto-merge only looked for an existing open PR and skipped if none was found. Now it auto-creates a PR from the dev branch to int-agentic if one doesn't exist, then merges it. This ensures every successful pipeline run promotes code to int-agentic regardless of whether a PR was manually created beforehand. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22330447642 * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22334055353 --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * Promote: int-agentic → main (production) (#420) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * Updating Durable Agent Implementation (#404) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * chore: reduce anomaly probability to 1% for controlled demo pace * Reduce anomaly probability to 1% for controlled demo pace (#406) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * chore: reduce anomaly probability to 1% for controlled demo pace --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * add mcp_agent_demo * Add hybrid MCP server (strict-schema + natural-language tools) and typed-contract workflow Scripts 6-8 for the MCP agent demo: - workflow_typed_contracts.py: 4-agent IT security incident response pipeline with Pydantic-enforced contracts at every boundary (SecurityAlert, ThreatAssessment, ImpactAnalysis, IncidentResponse) - mcp_server_hybrid.py: Single MCP endpoint exposing BOTH tool types: * Strict-schema: triage_alert, assess_threat, create_response (Pydantic) * Natural-language: ask_security_advisor, explain_for_customer (prose) * Shared session state across all tools - mcp_client_hybrid.py: 5-step incident flow using both tool types with explicit context passing between steps - README.md: Updated from 6 to 8 capabilities with Hybrid Mode architecture diagram and Quick Start sections 7-8 * Replace workflow_local_remote.py with simplified proxy agent (Script 3) - Delete workflow_local_remote.py (old Script 3 with extra LLM call) - Add workflow_proxy_agent.py: MCPProxyAgent (BaseAgent, no LLM) calls MCP tools directly via call_tool() zero local LLM overhead - Simplify MCPProxyAgent: extract common _call() method, module-level _last_text() helper, remove verbose static methods - Fix workflow_typed_contracts.py: add async with context managers - Update README: merge Script 9 into Script 3 position, update architecture diagram, Quick Start, How It Works sections * Clean up MCP demos: remove comparison commentary, drop typed-contracts - workflow_proxy_agent.py: remove 'traditional vs proxy' framing MCPProxyAgent is the standard approach for remote agent integration - Delete workflow_typed_contracts.py: not relevant to MCP integration - README: 7 scripts, renumber hybrid 6-7, remove typed-contract sections * Add LangGraph + MAF GroupChat cross-framework demo (Scripts 8-9) - mcp_server_langgraph.py: LangGraph ReAct agent exposed as MCP server on port 8003 with architecture tools (pattern eval, migration estimate, tech stack recommendation) - workflow_group_chat.py: MAF GroupChatBuilder orchestration with LLM planner routing between local BusinessStrategist (MAF) and remote TechnicalArchitect (LangGraph via MCP) - pyproject.toml: add langgraph, langchain-openai, langchain-core deps - README.md: update to 9 scripts, add cross-framework architecture diagram, Quick Start sections, and dependency table entries * MCP agent demo: rewrite README as MCP-vs-A2A thesis, delete proxy agent, finalize group chat with inline Planner * README: Mermaid diagrams, professional tone, remove LinkedIn teasing language * README: reframe around multi-framework interop problem, two design patterns (Agent-as-Tool + Agent Adapter), add conceptual architecture * README: fix Mermaid diagrams - use br tags instead of \n for line breaks * README: fix A2A comparison table - elicitation supported, structured schemas not * Upgrade agent-framework to 1.0.0rc1 and fastmcp to 3.0.2 Breaking changes migrated: - agent-framework 1.0.0rc1: ChatAgent->Agent, AgentThread->AgentSession, run_stream->run(stream=True), model->default_options(ChatOptions), WorkflowOutputEvent/AgentRunEvent/RequestInfoEvent->unified WorkflowEvent, MagenticBuilder now uses constructor kwargs, orchestrations moved to agent_framework_orchestrations package - fastmcp 3.0.2: removed mcp version pin (now pulled by fastmcp), no code changes needed (fully compatible) Files updated: - agentic_ai/agents/agent_framework/single_agent.py - agentic_ai/agents/agent_framework/multi_agent/handoff_multi_domain_agent.py - agentic_ai/agents/agent_framework/multi_agent/reflection_agent.py - agentic_ai/agents/agent_framework/multi_agent/magentic_group.py - agentic_ai/applications/pyproject.toml (agent-framework==1.0.0rc1) - agentic_ai/workflow/fraud_detection_durable/pyproject.toml (agent-framework==1.0.0rc1) - mcp/pyproject.toml (fastmcp==3.0.2) - tests/test_agent_framework_rc1_regression.py (51 regression tests, all passing) * Fix CI: regenerate requirements.txt with agent-framework-core==1.0.0rc1 The Docker build uses requirements.txt (not pyproject.toml/uv.lock), and it still had agent-framework-core==1.0.0b260130 pinned, which conflicts with agent-framework==1.0.0rc1 requiring agent-framework-core==1.0.0rc1. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22315210980 * Fix: use GH_PAT for auto-merge to trigger downstream workflows Pushes made by GITHUB_TOKEN don't trigger other workflows (GitHub Actions limitation to prevent infinite loops). This meant the auto-merge into int-agentic never triggered promote-to-main.yml. Now uses secrets.GH_PAT with fallback to GITHUB_TOKEN so: 1. If GH_PAT is configured: merge triggers promote-to-main.yml 2. If not: merge still works, but promotion PR must be created manually * Add workflow_dispatch trigger to promote-to-main Allows manual triggering when auto-trigger is missed (e.g. after GITHUB_TOKEN-based merges that don't fire downstream workflows). * Fix: Cosmos DB auth + remove phantom reflection_workflow_agent 1. MCP Cosmos DB: Replace AzureCliCredential (not available in containers) with ManagedIdentityCredential (when AZURE_CLIENT_ID is set) or DefaultAzureCredential as fallback. This fixes the 'Azure CLI not found on path' error in production Container Apps. 2. Backend: Remove reflection_workflow_agent from DEFAULT_AVAILABLE_AGENTS and descriptions - the module file doesn't exist, causing it to show as a broken option in the UI agent selector. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22321497108 * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment after successful auto-merge to int-agentic. Only dev environments are destroyed; production is retained. * Trigger rebuild: deploy Cosmos DB auth fix and agent list cleanup * Fix: iterate ResponseStream directly, not .updates ResponseStream is AsyncIterable - iterate with 'async for chunk in stream:' The .updates property returns a Sequence (list), not an async iterator. This caused TypeError: 'async for' requires __aiter__, got list. Not caught in local tests because unit tests mocked the streaming and the regression tests verified API signatures but didn't run live agents. * Fix: make integration test failures block the pipeline Removed continue-on-error: true from the pytest step so test failures actually fail the workflow. Previously, 5/7 tests could fail and the pipeline would still report success and proceed to auto-merge. * Fix: use SHA-tagged images for container deployment, not :latest Container Apps don't create a new revision when the image tag is unchanged. Using :latest meant the old container kept running even after a new image was pushed. Now passes github.sha as the image tag, which forces a new revision on every deploy. This was the root cause of the ResponseStream.updates bug reaching production - the integration tests ran against old containers. * Skip MCP localhost check in CI mode In CI, the evaluation talks to the backend via HTTP which connects to the deployed MCP container internally. The localhost:8000 check is only useful for local dev and was producing a confusing warning in CI logs. * Fix: auto-merge creates PR if none exists Previously auto-merge only looked for an existing open PR and skipped if none was found. Now it auto-creates a PR from the dev branch to int-agentic if one doesn't exist, then merges it. This ensures every successful pipeline run promotes code to int-agentic regardless of whether a PR was manually created beforehand. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22330447642 * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22334055353 * Migrate workshop to agent-framework==1.2.1 (incl. native HandoffBuilder) (#419) * Promote: int-agentic → main (production) (#414) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * Updating Durable Agent Implementation (#404) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * chore: reduce anomaly probability to 1% for controlled demo pace * Reduce anomaly probability to 1% for controlled demo pace (#406) * update fraud_detection_durable to feb 12 agent-framework * enhance fraud detection durable * update readme * chore: reduce anomaly probability to 1% for controlled demo pace --------- Co-authored-by: James N. <james.nguyen@microsoft.com> * add mcp_agent_demo * Add hybrid MCP server (strict-schema + natural-language tools) and typed-contract workflow Scripts 6-8 for the MCP agent demo: - workflow_typed_contracts.py: 4-agent IT security incident response pipeline with Pydantic-enforced contracts at every boundary (SecurityAlert, ThreatAssessment, ImpactAnalysis, IncidentResponse) - mcp_server_hybrid.py: Single MCP endpoint exposing BOTH tool types: * Strict-schema: triage_alert, assess_threat, create_response (Pydantic) * Natural-language: ask_security_advisor, explain_for_customer (prose) * Shared session state across all tools - mcp_client_hybrid.py: 5-step incident flow using both tool types with explicit context passing between steps - README.md: Updated from 6 to 8 capabilities with Hybrid Mode architecture diagram and Quick Start sections 7-8 * Replace workflow_local_remote.py with simplified proxy agent (Script 3) - Delete workflow_local_remote.py (old Script 3 with extra LLM call) - Add workflow_proxy_agent.py: MCPProxyAgent (BaseAgent, no LLM) calls MCP tools directly via call_tool() zero local LLM overhead - Simplify MCPProxyAgent: extract common _call() method, module-level _last_text() helper, remove verbose static methods - Fix workflow_typed_contracts.py: add async with context managers - Update README: merge Script 9 into Script 3 position, update architecture diagram, Quick Start, How It Works sections * Clean up MCP demos: remove comparison commentary, drop typed-contracts - workflow_proxy_agent.py: remove 'traditional vs proxy' framing MCPProxyAgent is the standard approach for remote agent integration - Delete workflow_typed_contracts.py: not relevant to MCP integration - README: 7 scripts, renumber hybrid 6-7, remove typed-contract sections * Add LangGraph + MAF GroupChat cross-framework demo (Scripts 8-9) - mcp_server_langgraph.py: LangGraph ReAct agent exposed as MCP server on port 8003 with architecture tools (pattern eval, migration estimate, tech stack recommendation) - workflow_group_chat.py: MAF GroupChatBuilder orchestration with LLM planner routing between local BusinessStrategist (MAF) and remote TechnicalArchitect (LangGraph via MCP) - pyproject.toml: add langgraph, langchain-openai, langchain-core deps - README.md: update to 9 scripts, add cross-framework architecture diagram, Quick Start sections, and dependency table entries * MCP agent demo: rewrite README as MCP-vs-A2A thesis, delete proxy agent, finalize group chat with inline Planner * README: Mermaid diagrams, professional tone, remove LinkedIn teasing language * README: reframe around multi-framework interop problem, two design patterns (Agent-as-Tool + Agent Adapter), add conceptual architecture * README: fix Mermaid diagrams - use br tags instead of \n for line breaks * README: fix A2A comparison table - elicitation supported, structured schemas not * Upgrade agent-framework to 1.0.0rc1 and fastmcp to 3.0.2 Breaking changes migrated: - agent-framework 1.0.0rc1: ChatAgent->Agent, AgentThread->AgentSession, run_stream->run(stream=True), model->default_options(ChatOptions), WorkflowOutputEvent/AgentRunEvent/RequestInfoEvent->unified WorkflowEvent, MagenticBuilder now uses constructor kwargs, orchestrations moved to agent_framework_orchestrations package - fastmcp 3.0.2: removed mcp version pin (now pulled by fastmcp), no code changes needed (fully compatible) Files updated: - agentic_ai/agents/agent_framework/single_agent.py - agentic_ai/agents/agent_framework/multi_agent/handoff_multi_domain_agent.py - agentic_ai/agents/agent_framework/multi_agent/reflection_agent.py - agentic_ai/agents/agent_framework/multi_agent/magentic_group.py - agentic_ai/applications/pyproject.toml (agent-framework==1.0.0rc1) - agentic_ai/workflow/fraud_detection_durable/pyproject.toml (agent-framework==1.0.0rc1) - mcp/pyproject.toml (fastmcp==3.0.2) - tests/test_agent_framework_rc1_regression.py (51 regression tests, all passing) * Fix CI: regenerate requirements.txt with agent-framework-core==1.0.0rc1 The Docker build uses requirements.txt (not pyproject.toml/uv.lock), and it still had agent-framework-core==1.0.0b260130 pinned, which conflicts with agent-framework==1.0.0rc1 requiring agent-framework-core==1.0.0rc1. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22315210980 * Fix: use GH_PAT for auto-merge to trigger downstream workflows Pushes made by GITHUB_TOKEN don't trigger other workflows (GitHub Actions limitation to prevent infinite loops). This meant the auto-merge into int-agentic never triggered promote-to-main.yml. Now uses secrets.GH_PAT with fallback to GITHUB_TOKEN so: 1. If GH_PAT is configured: merge triggers promote-to-main.yml 2. If not: merge still works, but promotion PR must be created manually * Add workflow_dispatch trigger to promote-to-main Allows manual triggering when auto-trigger is missed (e.g. after GITHUB_TOKEN-based merges that don't fire downstream workflows). * Fix: Cosmos DB auth + remove phantom reflection_workflow_agent 1. MCP Cosmos DB: Replace AzureCliCredential (not available in containers) with ManagedIdentityCredential (when AZURE_CLIENT_ID is set) or DefaultAzureCredential as fallback. This fixes the 'Azure CLI not found on path' error in production Container Apps. 2. Backend: Remove reflection_workflow_agent from DEFAULT_AVAILABLE_AGENTS and descriptions - the module file doesn't exist, causing it to show as a broken option in the UI agent selector. * chore: merge james-dev into int-agentic (auto) Auto-merged after successful CI/CD pipeline run 22321497108 * Fix promote-to-main + auto-destroy dev environments 1. promote-to-main.yml: Use GH_PAT instead of GITHUB_TOKEN so the workflow can create PRs (GITHUB_TOKEN lacks permission for this in repos with branch protection). 2. orchestrate.yml: Add Step 8 auto-destroy that tears down the integration-* environment…

james-tn and others added 7 commits February 23, 2026 18:28

Bump agent-framework to 1.2.1 and migrate AzureOpenAIChatClient → Ope…

d2b3a27

…nAIChatClient Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com>

Rewrite handoff agent on native HandoffBuilder; update regression tes…

db126c8

…ts for 1.2.1 Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com>

Update HANDOFF_README and dependency table for native HandoffBuilder …

0b53aca

…migration Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com>

Address GPT-5.5 review: fix CheckpointStorage protocol, reflection_ag…

64f3868

…ent kwargs, observability sample Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com>

Fix PEP 8 spacing in mcp_agent_demo files (code review nits)

9461c2d

Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/b1edc1aa-9c20-46c5-afdc-0256050a03de Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com>

Copilot AI assigned Copilot and james-tn Apr 28, 2026

Copilot created this pull request from a session on behalf of james-tn April 28, 2026 16:42 View session

Copilot AI requested a review from james-tn April 28, 2026 16:42

Copilot finished work on behalf of james-tn April 28, 2026 16:42

Copilot started work on behalf of james-tn April 28, 2026 16:47 View session

Copilot finished work on behalf of james-tn April 28, 2026 16:51

james-tn changed the base branch from main to int-agentic April 28, 2026 16:54

Copilot started work on behalf of james-tn April 28, 2026 17:09 View session

Regenerate uv.lock files; pass workflow_name to 1.2.x checkpoint listing

038dfb6

Agent-Logs-Url: https://github.com/microsoft/OpenAIWorkshop/sessions/91307ce7-2444-4b5f-b75f-2b736bc022c5 Co-authored-by: james-tn <25941658+james-tn@users.noreply.github.com>

james-tn approved these changes Apr 28, 2026

View reviewed changes

Copilot finished work on behalf of james-tn April 28, 2026 17:13

Copilot AI requested a review from james-tn April 28, 2026 17:13

Copilot AI had a problem deploying to integration April 28, 2026 17:13 Failure

james-tn marked this pull request as ready for review April 28, 2026 17:16

james-tn merged commit 8dc4983 into int-agentic Apr 28, 2026
12 of 13 checks passed

james-tn mentioned this pull request Apr 28, 2026

Fix backend 500s by renaming ChatOptions(model_id=…) to ChatOptions(model=…) for agent-framework 1.2.1 #421

Merged

9 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Migrate workshop to agent-framework==1.2.1 (incl. native HandoffBuilder)#419

Migrate workshop to agent-framework==1.2.1 (incl. native HandoffBuilder)#419
james-tn merged 8 commits into
int-agenticfrom
copilot/migrate-to-agent-framework-1-2-1

Copilot AI commented Apr 28, 2026

Uh oh!

james-tn commented Apr 28, 2026

Uh oh!

Copilot AI commented Apr 28, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

Copilot AI commented Apr 28, 2026

Investigation: what changes between 1.0.0rc1 and 1.2.1

What this PR does

Validation

Things to be aware of

Uh oh!

james-tn commented Apr 28, 2026

Uh oh!

Copilot AI commented Apr 28, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants