Stateful ReAct Agents Boost LLM Experimentation Efficiency.

Autonomous experimentation, often called the autoresearch pattern, allows large language models (LLMs) to iteratively refine code to optimize specific metrics. However, current stateless designs incur high token costs because they rebuild the experimental context from scratch in each iteration. This leads to a quadratic increase in total token usage over time. A new approach reformulates this pattern using a stateful ReAct agent, built with LangGraph. This agent leverages typed persistent state to carry experimental history across iterations via a tool-calling interface. This design fundamentally changes how context is managed, moving from an O(n) cost per iteration to a fixed-size conversation window with O(1) cost. Evaluations on hyperparameter tuning and code performance optimization benchmarks demonstrated significant token reductions. The stateful agent achieved 90% fewer tokens for hyperparameter tuning and 52% fewer for code optimization, all while maintaining comparable optimization quality. The paper provides architectural details for practitioners to implement this token-efficient method.