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.

loopgain.ai · Benchmarks · Dashboard · PyPI

Verify-revise loops — agentic coding, self-refinement, ReAct — have no real stopping signal, so they run to a guessed max_iterations. Set it too low and you cut the loop off before it's done. Set it too high and you burn tokens and wall-clock on iterations that aren't improving anything — or, worse, that quietly degrade a correct answer back into a broken one. max_iterations=N is just a guess, and the loop has no idea which iteration was its best.

LoopGain watches each loop's error trajectory and classifies it live into five named states — FAST_CONVERGE, CONVERGING, STALLING, OSCILLATING, DIVERGING — then acts on that signal:

• Stops the loop once it has converged, instead of running out the cap.
• Rolls back to the best-so-far iteration before a loop degrades a good result.
• Estimates remaining iterations live, exposed as lg.eta.

Under the hood it's a Barkhausen-criterion (Aβ) stability classifier — the same loop-gain test that decides whether any feedback system converges or oscillates, applied to an LLM agent loop instead of an amplifier. That's the how; the outcome is less spend and faster loops.

pip install loopgain

from loopgain import LoopGain

lg = LoopGain(target_error=0.1)          # wrap your existing loop — framework-agnostic

while lg.should_continue():              # stops on convergence, not a fixed cap
    errors = verifier.verify(output)
    lg.observe(errors, output=output)
    output = reviser.revise(output, errors)

result = lg.result
print(result.outcome)        # "converged" | "stalled" | "oscillating" | "diverged" | "max_iterations"
print(result.best_output)    # best-so-far iteration, automatically recovered

Measured across a public benchmark of 2,000 paired real-API trials (8,000 loop runs) against a fixed max_iterations=20 baseline, on loopgain v0.4.0:

Weighted judge preference 0.678 across 1,800 pairwise comparisons, and zero of six pre-registered kill criteria fired. Full protocol, raw data, and the cases where it doesn't help are public: loopgain-bench.

Scope, honestly: LoopGain proves a loop stopped moving and recovers the best iteration it saw — it does not by itself prove the loop stopped at the correct answer. It's a cost-and-stability control on the loop, not a correctness oracle.

Six first-class adapters plus the raw API — framework- and model-agnostic, never tied to one provider:

LangGraph · CrewAI · AutoGen · LangChain · OpenAI Agents SDK · Claude Agent SDK

• loopgain — the library. Apache-2.0, free, self-hostable. PyPI · source
• Hosted dashboard + telemetry — paid SaaS for fleet-wide loop observability and alerting. dashboard.loopgain.ai

Source is open; hosting and ops are paid. Run it entirely yourself, or let us run the dashboard.