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

SkillsVote proposes a governance framework for agent skills that helps long-horizon LLM agents curate, recommend, and evolve reusable skills.

The system uses structured search, trajectory analysis, and evidence-gated updates to rank skills and reduce noise from unverified additions.

By managing an external skill library instead of changing model weights, SkillsVote improves frozen-agent performance on benchmarks like Terminal-Bench 2.0 and SWE-Bench Pro.