SkillsVote Treats Agent Skills Like an Election
arXiv 2605.18401, 111 upvotes on HuggingFace Daily Papers. Hongyi Liu and co-authors propose a lifecycle governance framework for agent skill libraries — collection, recommendation, evolution — that gates which skills get into the library and which don't. Results: +7.9 points on Terminal-Bench 2.0, +2.6 on SWE-Bench Pro.
The "voting" framing is the interesting move. The last six months produced multiple skills frameworks (Anthropic Skills, Imbad0202 academic-research-skills, tech-leads-club agent-skills, scientific-agent-skills). All of them solve the "where do skills come from" problem. None really solved "how do you keep the library from poisoning itself with bad skills as agents add their own."
SkillsVote's answer: every skill update goes through evidence-gated approval — quality signals, environment requirements, source profile of the underlying code. The library doesn't accept a skill just because an agent wanted it. It accepts the skill if the evidence votes for it. Trajectories become candidates, candidates become skills only after the audit.
This pairs with the broader skills-security cluster — ClawHavoc, the New Stack vulnerability audit, OWASP Agentic Skills Top 10 draft, tech-leads-club's signed-skill registry. Provenance and content addressing was step one. Evidence-gated evolution is step two.
Paper: https://arxiv.org/abs/2605.18401
← Back to all articles
The "voting" framing is the interesting move. The last six months produced multiple skills frameworks (Anthropic Skills, Imbad0202 academic-research-skills, tech-leads-club agent-skills, scientific-agent-skills). All of them solve the "where do skills come from" problem. None really solved "how do you keep the library from poisoning itself with bad skills as agents add their own."
SkillsVote's answer: every skill update goes through evidence-gated approval — quality signals, environment requirements, source profile of the underlying code. The library doesn't accept a skill just because an agent wanted it. It accepts the skill if the evidence votes for it. Trajectories become candidates, candidates become skills only after the audit.
This pairs with the broader skills-security cluster — ClawHavoc, the New Stack vulnerability audit, OWASP Agentic Skills Top 10 draft, tech-leads-club's signed-skill registry. Provenance and content addressing was step one. Evidence-gated evolution is step two.
Paper: https://arxiv.org/abs/2605.18401
Comments