Can AI agents learn to reuse and improve their own skills?

LLM agents often solve tasks with ad-hoc skills that aren't reused or improved. MUSE-Autoskill treats skills as living assets with memory and testing, letting agents create skills on demand, store them for future tasks, and refine them via unit tests and runtime feedback. On SkillsBench, the approach improves task success, efficiency, and the ability to transfer skills between agents—suggesting that long-term skill management beats treating each task from scratch.