SkillEvo Improves LLM Agent Skills with Multi-Turn Interaction Feedback

Current methods for developing LLM agent skills often involve manual authoring or single-pass generation, lacking a continuous feedback loop for improvement based on actual interaction failures. While some approaches attempt to close this loop, they typically rely on single-turn evaluations, leading to a rapid decay in the evolution gradient once initial issues are resolved. This new research introduces SkillEvo, a framework designed to provide sustained skill evolution. SkillEvo generates trustworthy evolution gradients by transforming multi-turn user simulations into a continuous feedback mechanism, where follow-up questions progressively uncover deeper defects. It also incorporates a distinct governance layer that actively addresses factual inaccuracies and structural inefficiencies, preventing the degradation of skills over time.