SkillEvo: Self-Renewing Evolution Gradients from Multi-Turn Interaction Feedback

A recent submission on arXiv (2608.13120) presents SkillEvo, a novel approach for enhancing AI agent skills through feedback from multi-turn interactions. The authors highlight that existing agent skills are typically either manually created or produced in a single pass by a large language model (LLM), which fails to establish a feedback loop for improvement based on interaction errors. While recent advancements have attempted to close this loop, they depend on single-turn evaluations, resulting in a significant asymmetry: after initial adjustments, the evolution gradient diminishes, leaving multi-turn issues unaddressed. The governance of these systems relies on an end-to-end verification score, which can filter out poor candidates but lacks the ability to identify or rectify underlying issues. The paper suggests that the key limitation for ongoing skill evolution lies not in the number of iterations or editing capabilities, but in the reliability of evaluation feedback for providing effective evolution gradients. SkillEvo overcomes this challenge by leveraging multi-turn interaction feedback to refresh evolution gradients, facilitating continuous enhancement.