EvoSOP Lets LLM Agents Self-Evolve by Building Reusable Standard Operating Procedures

Current LLM agent frameworks rely on static, granular atomic actions (e.g., single-turn search, basic file I/O), forcing agents to repeatedly reinvent low-level logic for recurring workflows. This leads to high reasoning overhead and increased failure rates. To address this, the authors introduce EvoSOP, which enables agents to extract Standard Operating Procedures (SOPs) from execution trajectories. These SOPs act as higher-order tools that package multi-step logic into reusable, callable components, effectively allowing the agent to build and refine its own toolkit over time.

EvoSOP operates through a systematic lifecycle: construction (creating SOPs from trajectories), merging (combining similar SOPs), evaluation (testing performance), and pruning (removing ineffective SOPs). This iterative optimization fosters reliable and efficient tool-use patterns. The framework was tested on multiple benchmarks, demonstrating substantial improvements in task success rates and up to a 2x reduction in the number of interaction rounds. The paper provides a scalable pathway for developing truly self-evolving LLM agents that require less manual tool engineering.