LoopGain

An open-source cost controller for AI agent loops.

AI agent loops waste time and money when they don't know when to stop. LoopGain measures the loop in real time and stops it the moment it has actually converged — and rolls back before it degrades — instead of running to a fixed max_iterations cap.

Across 2,000 paired trials over 10 cells, LoopGain reduced total API spend by 92.8% vs max_iter=20, dropped median wall-clock latency from 30.9s to 2.1s (~15×), preserved output quality on natural-distribution workloads (W1–W4: judge winrate 0.50–0.63, CI excluding null on most cells), and improved output quality on engineered-failure workloads (W5: winrate 0.92–0.95 across three adapters). Weighted-average pairwise preference for LG vs B20 across 1,800 judge comparisons: 0.678. Zero of six kill criteria fired.

Home: loopgain.ai

Works for any iterative AI workflow with a measurable error signal — verify-revise loops, refinement passes, tool-use retry chains, RAG with self-correction, code-gen with linter feedback, multi-step reasoning loops. Pre-built adapters for LangGraph, CrewAI, AutoGen, LangChain, OpenAI Agents SDK, and Claude Agent SDK; drop-in via the raw API for any custom stack. Pure Python, no runtime dependencies.

Keywords: AI agent loops · agentic AI · infinite loop detection · divergence detection · early stopping · convergence · agent orchestration · LLM stability · generator-verifier-reviser · feedback-loop control.

---

Why

Production agent loops universally use max_iterations=N as their termination policy. It's the embarrassing default of agentic AI: you either waste compute (loop stops too late) or ship bad output (loop stops too early). LoopGain replaces it with a control-theoretic stop-and-rollback policy grounded in the Barkhausen criterion — a foundational result from electrical-engineering feedback-oscillator analysis (1921).

---

Install

pip install loopgain

Pure Python, no dependencies, supports Python 3.10+.

---

Usage

Three lines of code wrap any iterative loop with a measurable error signal:

from loopgain import LoopGain

lg = LoopGain(target_error=0.1)

while lg.should_continue():
    errors = verifier.verify(output)
    lg.observe(errors, output=output)
    output = reviser.revise(output, errors)

result = lg.result
print(result.outcome)              # "converged" | "oscillating" | "diverged" | "stalled" | "max_iterations"
print(result.best_output)          # the lowest-error iteration's output
print(result.iterations_used)
print(result.gain_margin)          # 1 / max(Aβ_smooth)
print(result.savings_vs_fixed_cap)

observe() accepts either a numeric error magnitude or any sequence (whose length becomes the magnitude). Pass output=... to enable the best-so-far buffer.

---

Defining your error signal

The one thing you provide is the error signal: a single non-negative number, every iteration, that says how wrong the current output is. Lower is better; zero means done. LoopGain doesn't know what your loop does — it just watches that number's trajectory and decides whether to keep going, stop, or roll back.

Your loop already has some way of knowing the output isn't good yet (or it wouldn't keep revising). Turn that into a number:

Loop	Error signal =
Agentic coding (write code → run tests)	number of failing tests (10 → 3 → 0)
JSON / structured extraction	number of schema violations
RAG with self-correction	number of required facts still missing
Self-refinement with an LLM judge	judge's gap to target (e.g. 10 − quality_score)
Lint / format loop	lint error count

The only rules: non-negative, and smaller as the output gets better. Returning the raw list of problems works directly — observe() uses its length as the magnitude (e.g. hand it the list of failing tests).

If your quality is fuzzy and has no natural "zero," run with target_error=None: LoopGain then stops when the number stops improving, wherever that plateau is, instead of waiting for an exact target.

Every stop/continue decision is made from this one number, so LoopGain is only as good as the error signal you give it — pick one that genuinely tracks output quality.

---

How it works

LoopGain measures empirical loop gain (Aβ = E(n) / E(n-1)) at every iteration and exposes it as a smoothed time series for visualization. The decision engine, however, classifies the full error trajectory using four features:

E_ratio   = E_current / E_first      # cumulative reduction
slope_log = OLS slope of log10(E)    # geometric trend direction
slope_p   = t-test p-value of slope  # statistical significance
osc_std   = std of detrended log10(E) # oscillation magnitude

It routes the trajectory into one of five named states:

State	Condition	Action
FAST_CONVERGE	cumulative reduction to ≤ 10% of E_first	Continue
CONVERGING	negative slope with p < 0.05, OR cumulative ≤ 50%	Continue, watch for upward drift
STALLING	no significant slope, no detectable oscillation	Stop after 2 consecutive readings — return best-so-far
OSCILLATING	high residual variance with flat trend	Stop — return best-so-far
DIVERGING	positive slope with p < 0.05 AND cumulative > 110%	Abort — roll back to best-so-far

Plus a short-circuit: if observed error drops at or below target_error, the loop stops immediately with state TARGET_MET. The default target_error=0.0 short-circuits on exactly zero error — the natural completion signal for verifier-driven loops. Pass target_error=None to disable the short-circuit and rely on stability detection alone.

The decision is conservative by design: requiring both statistical significance and meaningful cumulative motion before terminating prevents false-positive aborts on noisy real-LLM error series. Validated at 98.8% macro-averaged accuracy across 5 regimes on N=1000 deterministic-mock trajectories (see RESULTS_v2_classifier.md). The STALLING ceiling of ~94% is the t-test's irreducible 5% type-I error rate, not a classifier weakness.

Recommended minimum: 6 iterations for reliable trend significance. At n≤4 the t-test is severely underpowered (df=2 requires |t|>4.3 for p<0.05) — the classifier conservatively falls back to STALLING when evidence is thin. The thresholds are derived analytically (control theory + statistical convention), not fitted; tune them per domain via the TrajectoryThresholds argument once you have production traces.

Legacy single-feature classifier: the original v0.1 single-Aβ-band classifier (thresholds 0.3 / 0.85 / 0.95 / 1.05) is still available via LoopGain(classifier='legacy_bands') for callers that have empirically tuned the bands to a specific workload.

---

Best-so-far rollback

LoopGain keeps a buffer of all observed outputs paired with their error scores. On termination it returns argmin(error), not the last iteration:

Terminal state	Returned output
TARGET_MET	Current output (by definition, the best)
OSCILLATING	Lowest-error iteration in the buffer
DIVERGING	Lowest-error iteration (which is not the last one)

This transforms divergence detection from "abort with garbage" into "abort with the best you've seen so far" — a free quality floor.

---

What LoopGain does and doesn't guarantee

LoopGain saves money by stopping a loop once it stops improving — fewer iterations, fewer tokens. In our public benchmark, that was a 92.8% median cut in API spend vs max_iterations=20, with output quality preserved. Two honest limits:

• Savings depend on your workload. Loops that usually succeed fast save the most (~96%); adversarial, failure-prone loops save less (~78–84%). The headline is a blend — run the benchmark on your own loops before quoting a number.
• LoopGain detects convergence, not correctness. It stops when your error signal stops improving — which means more iterations won't help, not that the loop succeeded. On the benchmark this preserved quality (it rarely stopped early on a worse output; false-stop rate ≤4.5%), but a loop can stall with the error still above zero — a plateau at, say, 2 failing tests. So check result.best_error (or your own pass/fail) before you trust the output: if it plateaued short of your target, that's a quality gap LoopGain can't see, and a false stop that forces a rerun is the one way it eats into the savings. LoopGain decides when to stop; you decide whether the answer is good enough.

---

API reference

LoopGain(target_error=0.0, max_iterations=50, thresholds=None, trajectory_thresholds=None, classifier='trajectory', smoothing_window=3, assumed_fixed_cap=10)

Construct the monitor.

• target_error — Stop when an observed error drops at or below this. Default 0.0 short-circuits on exactly zero error (the natural completion signal for verifier-driven loops). Pass None to disable the short-circuit entirely.
• max_iterations — Hard safety backstop. Default 50 so the loop can never run unbounded; a stability verdict normally terminates it well before this. Pass None to opt into a fully unbounded loop (only safe if your loop is guaranteed to reach target_error or a stop-state), or a smaller integer to cap tighter.
• thresholds — Custom ThresholdBands for the legacy single-Aβ-band classifier. Ignored when classifier='trajectory'.
• trajectory_thresholds — Custom TrajectoryThresholds for the multi-feature classifier (the default). Override only with workload-specific evidence.
• classifier — 'trajectory' (default, v0.2 multi-feature classifier) or 'legacy_bands' (v0.1 single-Aβ-band classifier).
• smoothing_window — EMA window for the smoothed Aβ series (always maintained for visualization, regardless of classifier choice). Default 3.
• assumed_fixed_cap — Used to compute savings_vs_fixed_cap. Default 10.

lg.observe(errors, output=None) -> str

Record this iteration's errors and optional output. Returns the current state name. errors accepts a number (used directly) or any sequence (length used as magnitude).

lg.should_continue() -> bool

Returns False once a terminal state fires.

lg.state -> str

Current state name. One of INIT, FAST_CONVERGE, CONVERGING, STALLING, OSCILLATING, DIVERGING, TARGET_MET, MAX_ITERATIONS. The corresponding terminal result.outcome values are converged, oscillating, diverged, stalled (v0.2 trajectory mode only — STALLING terminating after 2 consecutive readings), max_iterations, or in_progress.

lg.eta -> int | None

Best-effort closed-form estimate of iterations remaining, exposed for instrumentation. Returns None whenever it isn't well-defined — which is most of the time on real, jump-dominated loops, so don't depend on it for control.

lg.gain_margin -> float | None

1 / max(Aβ_smooth). > 1 means stable headroom across the entire run.

lg.result -> LoopGainResult

Terminal result with outcome, iterations_used, best_index, best_output, best_error, convergence_profile, error_history, gain_margin, savings_vs_fixed_cap. Safe to call mid-loop.

lg.send_telemetry(endpoint, token, workload_id=None, timeout=2.0, allow_insecure=False, framework=None, loop_type=None, team=None, include_per_iteration=True) -> bool

Opt-in. Send a single anonymized telemetry POST after the loop terminates. Best-effort — never raises, returns True on 2xx, False otherwise. Adapters auto-stamp framework; loop_type and team are free-form labels that surface as filters in the dashboard. Pass include_per_iteration=False to send aggregate summary only.

import os
from loopgain import LoopGain

lg = LoopGain(target_error=0.1)
# ... run the loop ...
lg.send_telemetry(
    endpoint=os.environ["LOOPGAIN_TELEMETRY_ENDPOINT"],   # or hardcode
    token=os.environ["LOOPGAIN_TELEMETRY_TOKEN"],         # never hardcode
    workload_id="my-rag-pipeline",                        # opaque label
)

Recommended setup: store the token outside source. Two clean options:

# Option A: environment variable (simplest)
export LOOPGAIN_TELEMETRY_ENDPOINT="https://telemetry.loopgain.ai/v1/aggregate"
export LOOPGAIN_TELEMETRY_TOKEN="lgk_..."   # add to ~/.zshrc or ~/.bashrc

# Option B: macOS Keychain (more secure)
pip install keyring
python3 -c "import keyring; keyring.set_password('loopgain', 'telemetry', input('Token: '))"
# Then in code: keyring.get_password('loopgain', 'telemetry')

What is sent: state transitions, Aβ summary (min/max/median), gain margin, rollback flag, iterations used, savings, library version, optional opaque workload_id, threshold config, hour-bucketed timestamp.

What is NEVER sent: prompts, completions, error contents, output buffer, individual Aβ values, or any customer identity beyond the bearer token. Privacy contract is enforced by the payload-shape unit tests in tests/test_telemetry.py.

The hosted endpoint at telemetry.loopgain.ai is one acceptable destination. The receiver and dashboard are both open-source — self-host to keep telemetry fully under your control.

This is not the same as anonymous usage telemetry. send_telemetry sends your loop data to your dashboard, and only when you call it. There's a separate, opt-in funnel telemetry described below. The two never share data or code.

---

Anonymous funnel telemetry (opt-in, off by default)

LoopGain can report anonymous usage counts so a solo maintainer can tell whether the library is actually being used — install → first observe() → recurring use. It is opt-in and default-decline: nothing is sent unless you explicitly turn it on.

loopgain telemetry --show       # status + exactly what would be sent
loopgain telemetry --enable     # opt in   (or: export LOOPGAIN_TELEMETRY=1)
loopgain telemetry --disable    # opt out  (or: export LOOPGAIN_TELEMETRY=0)

DO_NOT_TRACK=1 is honored as a hard opt-out, and CI environments are auto-detected and declined silently. When enabled, payloads carry only a locally-generated random id (not derived from your machine), hour-bucketed timestamps, library/Python/OS versions, the adapter in use, and a coarse outcome count. Prompts, outputs, error contents, keys, paths, and IPs are never collected. Delivery is batched, async, https-only, and fail-silent — it can never break your loop. Full details and the privacy contract: TELEMETRY.md.

---

Command-line interface

loopgain --version              # or: loopgain version
loopgain telemetry --show       # inspect / control anonymous funnel telemetry
python -m loopgain telemetry --show   # equivalent, without the console script

---

Framework adapters

Thin wrappers under loopgain.integrations drive each major agent framework's iteration with a LoopGain monitor and auto-stamp framework="<name>" on telemetry. The frameworks themselves are optional dependencies — install the extra you need:

pip install 'loopgain[langgraph]'          # LangGraph
pip install 'loopgain[crewai]'             # CrewAI
pip install 'loopgain[autogen]'            # AutoGen v0.4+
pip install 'loopgain[langchain]'          # LangChain (create_agent / AgentExecutor)
pip install 'loopgain[openai-agents]'      # OpenAI Agents SDK
pip install 'loopgain[claude-agent-sdk]'   # Anthropic Claude Agent SDK
pip install 'loopgain[all]'                # all six

All adapters take a LoopGain instance plus an error_fn you provide — the framework doesn't know what your error signal is, so the adapter doesn't either. error_fn returns a non-negative number (or None to skip an iteration).

LangGraph

Drives graph.stream(input, stream_mode="updates"). Each update is one iteration.

from loopgain import LoopGain
from loopgain.integrations import LangGraphAdapter

graph = build_my_verify_revise_graph().compile()
lg = LoopGain(target_error=0.1, max_iterations=20)

adapter = LangGraphAdapter(
    lg=lg,
    error_fn=lambda update: len(update.get("verifier", {}).get("errors", [])),
)
final_state = adapter.run(graph, {"draft": initial})

lg.send_telemetry(
    endpoint=os.environ["LOOPGAIN_TELEMETRY_ENDPOINT"],
    token=os.environ["LOOPGAIN_TELEMETRY_TOKEN"],
    workload_id="rag-rewrite",
    framework=adapter.framework_name,        # "langgraph", auto-stamped
)

adapter.stream(...) yields each item if you want the full trace; adapter.arun(...) / adapter.astream(...) are the async counterparts and accept an async error_fn.

CrewAI

Installs step_callback and/or task_callback on a Crew. Pick whichever granularity matches your loop — step_error_fn for refinement within a Task, task_error_fn for refinement across Tasks.

from crewai import Crew
from loopgain import LoopGain
from loopgain.integrations import CrewAIAdapter

lg = LoopGain(target_error=0.1, max_iterations=20)
adapter = CrewAIAdapter(
    lg=lg,
    task_error_fn=lambda task_output: count_failed_checks(task_output.raw),
)
crew = Crew(agents=[...], tasks=[...])
adapter.install(crew)
result = crew.kickoff()
adapter.uninstall()         # or use `with CrewAIAdapter(...) as a:` context

lg.send_telemetry(
    endpoint=...,
    token=...,
    framework=adapter.framework_name,        # "crewai"
)

The adapter chains with any callback you already had installed — your existing instrumentation isn't overwritten.

AutoGen (v0.4+)

Wraps team.run_stream(task=...). In a verify-revise rotation, filter to the verifier's messages with observe_sources={"verifier"} so only it drives observe().

from autogen_agentchat.teams import RoundRobinGroupChat
from loopgain import LoopGain
from loopgain.integrations import AutoGenAdapter

team = RoundRobinGroupChat(participants=[generator, verifier])
lg = LoopGain(target_error=0.1, max_iterations=20)
adapter = AutoGenAdapter(
    lg=lg,
    error_fn=lambda msg: parse_verifier_score(msg.content),
    observe_sources={"verifier"},
)
result = await adapter.run(team, task="...")

lg.send_telemetry(
    endpoint=...,
    token=...,
    framework=adapter.framework_name,        # "autogen"
)

Pass a cancellation_token to adapter.run(...) and the adapter will cancel it when LoopGain reaches a terminal state (target met, oscillation, divergence). The legacy v0.2 ConversableAgent.initiate_chat API is not supported — use the v0.4 event-driven runtime.

LangChain

Duck-types against any LangChain agent that exposes .stream(input, **kwargs) / .astream(input, **kwargs) — both the current langchain.agents.create_agent() (v1+) and the legacy AgentExecutor. The adapter forwards **stream_kwargs verbatim, so the chunk shape your error_fn sees is the one your agent emits.

from langchain.agents import create_agent
from loopgain import LoopGain
from loopgain.integrations import LangChainAdapter

agent = create_agent(model="gpt-5-nano", tools=[get_weather])
lg = LoopGain(target_error=0.0, max_iterations=20)

def error_fn(chunk):
    if chunk.get("type") != "updates":
        return None
    # Count unresolved tool calls; drops to 0 once the agent stops calling tools.
    return sum(
        1 for _, update in chunk["data"].items()
        if getattr(update.get("messages", [None])[-1], "tool_calls", None)
    )

adapter = LangChainAdapter(lg=lg, error_fn=error_fn)
final = adapter.run(
    agent,
    {"messages": [{"role": "user", "content": "What's the weather?"}]},
    stream_mode="updates",
    version="v2",
)

lg.send_telemetry(
    endpoint=...,
    token=...,
    framework=adapter.framework_name,        # "langchain"
)

For legacy AgentExecutor: just drop the stream_mode / version kwargs; each yielded chunk is an AddableDict per step (parse intermediate_steps or the terminal output key in your error_fn).

OpenAI Agents SDK

Wraps Runner.run_streamed(agent, input).stream_events(). The SDK is async-first; the adapter mirrors that. A run_sync helper wraps the async path with asyncio.run for synchronous callers.

from agents import Agent, function_tool
from loopgain import LoopGain
from loopgain.integrations import OpenAIAgentsAdapter

agent = Agent(name="Reviser", instructions="...", tools=[...])

lg = LoopGain(target_error=0.0, max_iterations=20)

def error_fn(event):
    # Default observes only run_item_stream_event; pull the verifier's
    # reported failure count off tool outputs.
    if event.item.type == "tool_call_output_item":
        return float(event.item.output.get("failures", 0))
    return None

adapter = OpenAIAgentsAdapter(lg=lg, error_fn=error_fn)
result = await adapter.run(agent, input="Fix the bug.")
print(result.final_output)

lg.send_telemetry(
    endpoint=...,
    token=...,
    framework=adapter.framework_name,        # "openai-agents"
)

By default the adapter only forwards run_item_stream_event to error_fn — pass observe_event_types=None to see every event (including raw token deltas and agent-handoff notifications). When LoopGain reaches a terminal state, the adapter best-effort calls .cancel() on the underlying RunResultStreaming.

Claude Agent SDK

Wraps Anthropic's claude_agent_sdk.query(prompt=..., options=...) async iterator. By default observes only AssistantMessage (skips UserMessage / SystemMessage / ResultMessage); override with observe_message_types=None or a custom tuple.

from claude_agent_sdk import ClaudeAgentOptions, TextBlock
from loopgain import LoopGain
from loopgain.integrations import ClaudeAgentSDKAdapter

def error_fn(message):
    # Count `FAIL:` markers a self-verifying persona emits.
    for block in getattr(message, "content", []):
        if isinstance(block, TextBlock):
            return float(block.text.count("FAIL:"))
    return None

lg = LoopGain(target_error=0.0, max_iterations=20)
adapter = ClaudeAgentSDKAdapter(lg=lg, error_fn=error_fn)

options = ClaudeAgentOptions(system_prompt="Self-verify each draft.")
result = await adapter.run(
    prompt="Write a haiku about feedback loops.",
    options=options,
)

lg.send_telemetry(
    endpoint=...,
    token=...,
    framework=adapter.framework_name,        # "claude-agent-sdk"
)

For the bidirectional ClaudeSDKClient use case, pass message_iterator=client.receive_messages() instead of prompt=....

Custom integrations

For frameworks without an adapter, the raw LoopGain.observe() API works against any iterable. The adapters are 100-200 lines each — copy one of loopgain/integrations/{langgraph,crewai,autogen,langchain,openai_agents,claude_agent_sdk}.py as a starting point.

---

Status

Initial public release. Core library shipped (current version: see the PyPI badge at the top). Framework adapters (LangGraph, CrewAI, AutoGen, LangChain, OpenAI Agents SDK, Claude Agent SDK) are installable as optional extras. The cloud-aggregator telemetry receiver and dashboard are live as separate open-source repos. The math and the API surface are stable.

This is alpha software. The API may break before 1.0 if production usage surfaces design issues; pin the version.

---

License

Apache-2.0.

---

Background

LoopGain applies the Barkhausen stability criterion (Heinrich Barkhausen, 1921 — the foundational result on when feedback amplifiers oscillate) to AI agent feedback loops. The criterion was originally a way to predict whether an electronic oscillator would sustain oscillation; it turns out to map cleanly onto any feedback loop you can attach an error signal to.

The cleanest summary: an iterative AI loop with a measurable error signal is a feedback system. The ratio E(n) / E(n-1) is its empirical loop gain. The Barkhausen result tells you that loop gain less than 1 converges, equal to 1 oscillates, greater than 1 diverges. LoopGain operationalizes this: classifies the loop's current band, decides what to do, and tells you when you'll converge.

Loop types this applies to in practice:

• Verify-revise loops (GVR pattern) — generator produces, verifier finds issues, reviser fixes. Error = issue count or severity-weighted score.
• Refinement loops — initial output, iterate to improve. Error = distance from target spec / rubric score.
• Tool-use retry chains — agent calls tool, gets back error/success, retries. Error = consecutive failure count or aggregate score.
• RAG with self-correction — retrieve, generate, critique, re-retrieve. Error = critique severity or hallucination score.
• Code generation with linter/test feedback — generate, run tests/linter, fix, repeat. Error = failing test count or linter violation count.
• Multi-step reasoning loops — ReAct-style think/act/observe iterations. Error = whatever the agent's quality assessor returns.
• Custom feedback loops — anything where you can produce a number that should drop toward zero as the loop succeeds.