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

Researchers propose MUSE-Autoskill, a framework enabling LLM agents to autonomously create, store, and refine reusable skills throughout their operational lifecycle. The system treats skills as long-lived, testable assets with integrated memory and evaluation mechanisms, demonstrating improved task success rates and cross-agent knowledge transfer on benchmark tests.

MUSE-Autoskill addresses a fundamental limitation in current LLM agent architectures: the static nature of task-solving capabilities. Traditional approaches treat skills as isolated components, preventing agents from learning from past experiences or efficiently transferring knowledge across tasks. This research introduces a lifecycle-managed approach where skills accumulate experience data, adapt to new contexts, and improve through continuous evaluation. The framework's integration of skill-level memory enables agents to build institutional knowledge, similar to how humans refine expertise over time.

The significance lies in moving beyond prompt engineering and toward genuine agent autonomy. By enabling self-evolution through structured skill creation and refinement, the framework addresses scalability challenges that plague current AI systems. Existing LLM agents struggle with long-horizon problem-solving partly because they lack mechanisms to consolidate and reuse learnings systematically. MUSE-Autoskill's unit testing and runtime feedback mechanisms provide quantifiable improvement pathways.

For the AI development community, this framework could accelerate progress toward more capable autonomous systems by reducing the engineering overhead required to maintain and improve agent capabilities. The cross-agent transfer capability suggests potential for distributed learning paradigms where multiple agents benefit from collectively refined skills. This has implications for enterprise AI deployment, where skill libraries could become valuable digital assets. Developers building multi-agent systems could leverage shared skill repositories to reduce redundancy and improve reliability across applications.

• →MUSE-Autoskill enables LLM agents to autonomously create, manage, and refine reusable skills through an integrated lifecycle framework.
• →Skill-level memory allows agents to accumulate experience across tasks, improving reusability and enabling more effective adaptation over time.
• →The framework demonstrates improved task success rates, efficiency gains, and successful cross-agent skill transfer on SkillsBench benchmarks.
• →Integration of unit testing and runtime feedback mechanisms provides quantifiable evaluation and continuous refinement pathways for agent capabilities.
• →This approach treats skills as long-lived, testable assets rather than static artifacts, representing a structural shift in agent architecture design.