Ziyu Ma et al. · 2026-04-09 · notable
SkillClaw: Let Skills Evolve Collectively
Framework that lets LLM-agent skill libraries improve from the trajectories of many users at once — an 'autonomous evolver' mines recurring patterns and rewrites the shared skill set without extra human work.
A cross-user, cross-session evolver that turns everyone's agent runs into a steadily-improving shared skill library.
What is it?
SkillClaw is a new framework from a team led by Ziyu Ma that attacks a specific gap in deployed LLM agents: most skill libraries are static, and every user's sessions are thrown away after they finish. SkillClaw treats those sessions as a collective training signal. It was the #1 Hugging Face paper of the day on April 10.
How does it work?
As users run an agent, SkillClaw continuously aggregates their trajectories and feeds them to an autonomous 'evolver' process. The evolver identifies recurring behavioral patterns across users and translates them into concrete updates to the shared skill repository — merging redundant skills, splitting overloaded ones, and promoting patterns that solved tasks. The agent then runs with the updated skill set on the next invocation, so improvements discovered by one user propagate system-wide with no manual curation.
Why does it matter?
This is a concrete, paper-backed recipe for the thing every agent product wants: 'my agent gets better the more people use it.' The key move is pushing skill evolution out of the model weights and into a cheap, auditable library update loop — which means teams can ship it into an existing agent without retraining.
Who is it for?
Agent framework authors, production agent teams shipping skill libraries.