LLM agents use static skills. but you can evolve them over time. this paper collects agent interactions, groups them by skills, an "agent evolver" analyzes them, then refines skills or creates new ones. in just 6 days with 8 users, it improves skills from 11% to 88%.