SkillsVote: Lifecycle Governance of Agent Skills from Collection, Recommendation to Evolution

SkillsVote is a framework for governing Agent Skills across collection, recommendation, attribution, and evolution.

It profiles a corpus of over one million open-source skills for environment requirements, quality, and verifiability.

Before task execution, it performs agentic library search to recommend relevant skills and generate usage guidance.

After execution, it decomposes trajectories into subtasks, attributes outcomes to skill-guided execution, agent exploration, environment, or result signals, and only admits successful, reusable discoveries into evidence-gated library updates.

Experiments on Terminal-Bench 2.0 and SWE-Bench Pro show consistent improvements over baselines without skills, with gains from online evolution and offline transfer via frozen libraries.

SkillsVote also outperforms SkillRouter in routing over large skill libraries, and recommendation reduces negative transfer by filtering harmful skill exposure.

The framework addresses the risks of irrelevant skills distracting agents and weakly supported experience polluting the library, positioning governed skill libraries as a practical substrate for scalable agent experience reuse.

Key results include up to +7.8 pp on Terminal-Bench 2.0 and +2.7 pp on SWE-Bench Pro, with routing FC@10 of 62.7% at full scale versus 46.7% for SkillRouter.

The system's attribution layer ensures conservative evolution, editing or creating skills only from attributed, reusable evidence.

Overall, SkillsVote demonstrates that lifecycle governance of skills improves agent performance and enables safe, scalable experience reuse.

The paper's contributions include the lifecycle framework, the million-scale corpus, the attribution-controlled evolution loop, and empirical validation on challenging benchmarks.

The approach is designed to be auditable and conservative, preventing spurious successes from being rewarded and keeping environment-caused failures from…