LoopGain Cuts AI Agent Loop Costs by 92%

LoopGain is a pure Python open-source library that, instead of a fixed cap like max_iterations=N, judges in real time whether an error signal has converged and automatically terminates or rolls back the loop.

How it works: Inspired by the Barkhausen criterion from electrical engineering, it measures an empirical loop gain (Aβ = E(n)/E(n-1)) at every iteration and classifies the error trajectory using 4 features (cumulative reduction ratio, OLS slope, t-test p-value, oscillation standard deviation).

There are 5 states:

• FAST_CONVERGE / CONVERGING → continue
• STALLING / OSCILLATING → stop, return the best value
• DIVERGING → abort, roll back to the best value

Usage takes just 3 lines. Initialize with LoopGain(target_error=0.1), then call lg.observe(errors, output=output) inside the loop, and that's it. If the target value is unclear, setting target_error=None makes it automatically terminate at the plateau point.

Benchmark (2,000 paired trials) results:

• 92.8% reduction in API spend compared to max_iter=20 ($27.05 → $1.94)
• Median processing time reduced from 30.9s to 2.1s (about 15x faster)
• Quality-weighted preference of 0.678 confirming quality is maintained

It ships with built-in adapters for LangGraph, CrewAI, AutoGen, LangChain, OpenAI Agents SDK, and Claude Agent SDK, with no runtime dependencies.