Self-Evolving LLM Agents Optimize Tools with SOPs

Large Language Model agents leverage tools to interact with the real world and tackle complex tasks. However, current agent frameworks often rely on static sets of basic, "atomic" actions, forcing agents to repeatedly reconstruct low-level logic for common workflows. This repetitive process leads to increased reasoning overhead and a higher likelihood of failure. This research introduces a novel approach where agents can "self-evolve" by creating and optimizing their own toolsets. The core idea is to synthesize these atomic actions into reusable "Standard Operating Procedures" (SOPs). These SOPs act as higher-order tools, encapsulating multi-step logic and allowing agents to execute complex sequences more efficiently. The EvoSOP framework facilitates this self-evolution through a systematic lifecycle: agents extract SOPs from their successful execution trajectories, merge similar procedures, evaluate their effectiveness, and prune less useful ones. Experiments demonstrate that EvoSOP significantly improves task success rates and drastically reduces the number of interaction rounds compared to traditional methods, fostering more reliable and efficient tool-use patterns for scalable agent development.