[FEATURE] Optional / Conditional Structured Output Support in structured_output_model

[pradeepdev-1995]

Problem Statement

Currently, when structured_output_model is enabled in Strands agents, the model is forced to always generate structured responses for every interaction, including simple conversational inputs such as:

"hi"
"hello"
"good morning"
"how are you"

This creates a poor conversational UX because small-talk responses are unnecessarily wrapped into structured JSON outputs.

Example:

Instead of:

Hello! How can I help you today?
the agent is forced to generate:

{
  "type": "smalltalk",
  "message": "Hello! How can I help you today?"
}

In real-world agentic applications, structured outputs are typically needed only for task-oriented requests such as:

extraction
workflows
API calls
automation
database operations
forms
business actions

but not for casual conversation.

The current SDK behavior applies structured_output_model globally for the entire invocation.

There is no official way to:

• dynamically disable structured output
• allow raw/plain-text responses for conversational queries
• conditionally switch between structured and unstructured responses
• support hybrid conversational + workflow agents cleanly

Proposed Solution

I also explored using:

Union[StructuredModel, SmallTalkModel]

and discriminated unions as a workaround.

However, the current implementation/tool schema conversion appears to flatten or degrade complex union schemas, leading to validation/tool-calling failures.

This suggests that advanced schema composition (Union, oneOf, polymorphic schemas) may not yet be fully supported in the structured output pipeline.

Use Case

Why This Matters

This feature is important for production-grade conversational agents because most real-world agents require:

natural conversation UX
workflow automation
structured extraction
tool invocation
mixed conversational behavior

within the same agent.

Currently developers must manually implement routing layers outside the SDK to achieve this cleanly.

Native support would significantly improve developer experience and make hybrid conversational/workflow agents easier to build using Strands.

Alternatives Solutions

Option 1 — structured_output_mode

Agent(
    structured_output_model=MySchema,
    structured_output_mode="auto"
)

Behavior:

task-oriented queries → structured output
casual/small-talk queries → plain text output

Option 2 — Native Hybrid Response Support

Allow:

Union[BaseModel, str]
or equivalent hybrid response handling officially.

Additional Context

No response