LangGraph Stateful ReAct Agent slashes token costs 90% in AI experiments

A new paper from Faramarz Jabbarvaziri on arXiv (2606.14945) tackles a major inefficiency in autonomous AI experimentation: the 'autoresearch' pattern, where LLMs iteratively modify code to optimize metrics. Traditional stateless agents reconstruct the full experimental context from scratch at every iteration, leading to O(n) token cost per step and O(n²) total. The proposed solution reformulates this as a stateful ReAct agent using LangGraph, where a typed persistent state carries experimental history across iterations via a tool-calling interface.

Benchmarked on two tasks—hyperparameter tuning (15 iterations, small observations) and code performance optimization (40 iterations, large source code and benchmark results)—the stateful agent achieves dramatic token reductions: 90% fewer on hyperparameter tuning (2,492 vs. 24,465 tokens) and 52% fewer on code optimization (627K vs. 1.275M tokens) while preserving optimization quality. The key structural improvement: stateless agents re-read full history with O(n) cost per iteration, while the stateful agent operates within a fixed-size conversation window at O(1) cost.

The paper provides enough architectural detail for practitioners to implement their own stateful autoresearch agents. This approach could significantly reduce API costs and latency for any workflow involving iterative LLM-driven experimentation, from hyperparameter tuning to automated code optimization. The practical impact is immediate for teams running large-scale experiment loops.