💡 The paper introduces SkillClaw, a framework that enables collective skill evolution in multi-user large language model agent systems. The problem addressed is that current systems rely on reusable skills that remain static after deployment, preventing the system from improving with experience. As a result, similar workflows and failure modes are repeatedly rediscovered across users. SkillClaw solves this issue by aggregating user interactions to autonomously update and improve skills across the ecosystem.

The method used in SkillClaw is to continuously aggregate trajectories generated during use and process them with an autonomous evolver. This evolver identifies recurring behavioral patterns and translates them into updates to the skill set by refining existing skills or extending them with new capabilities. The resulting skills are maintained in a shared repository and synchronized across users, allowing improvements discovered in one context to propagate system-wide without requiring additional effort from users.

The results of the paper show that SkillClaw significantly improves the performance of large language model agents in real-world scenarios, even with limited interaction and feedback. Experiments on WildClawBench demonstrate the effectiveness of SkillClaw in enabling cross-user knowledge transfer and cumulative capability improvement. Overall, the paper contributes a framework for collective skill evolution that enables large language model agent systems to learn from user interactions and improve over time.