MUSE-Autoskill: Self-Evolving Agents via Skill Creation, Memory, Management, and Evaluation

Researchers introduced MUSE-Autoskill, a skill-centric framework for LLM agents to create, store, manage, evaluate, and refine skills over time.

The system adds skill-level memory and uses unit tests plus runtime feedback to improve reuse, adaptation, and maintenance.

Early results on SkillsBench show gains in task success, efficiency, skill reuse, and transfer across agents.

The approach treats skills as reusable, testable assets, aiming to make agents more reliable and adaptable over the long term.