From Atomic Actions to Standard Operating Procedures: Iterative Tool Optimization for Self-Evolving LLM Agents

Existing LLM agent frameworks rely on static toolsets composed of granular atomic actions (e.g., basic file I/O or single-turn search), forcing agents to reinvent low-level logic for every recurring workflow, leading to increased reasoning overhead and failure rates. This work proposes EvoSOP, a framework that enables agents to synthesize atomic actions into callable higher-order tools called Standard Operating Procedures (SOPs). EvoSOP achieves iterative optimization through a four-stage lifecycle: extracting SOPs from execution trajectories, merging redundant SOPs, evaluating SOP utility, and pruning ineffective SOPs. Extensive experiments across various complex tasks show that EvoSOP not only significantly boosts task success rates but also substantially reduces the number of interaction rounds with the environment. Analysis reveals that iterative tool optimization fosters reliable and efficient tool-use patterns, providing a scalable pathway for self-evolving agents.