MUSE-Autoskill: A Skill Lifecycle Framework for Self-Evolving AI Agents

The research team proposes the MUSE-Autoskill framework, which allows large language model (LLM) agents to continuously accumulate and evolve skills through a unified lifecycle of five phases—creation, memory, management, evaluation, and optimization—achieving cross-task reuse and long-term improvement. The core innovation of this framework lies in treating skills as dynamic assets with a lifecycle and introducing a skill-level memory mechanism, opening up new possibilities for the self-evolution of agents.

LLM agents have evolved from question-answering systems to complex task performers, where skills—as reusable capability units—are the core. However, existing skill creation methods have three major limitations: isolation (lack of connections between skills, making experience transfer difficult), staticity (fixed skills that easily become outdated), and lack of evaluation mechanisms (uneven quality), which limit the reusability and long-term value of skills.

The MUSE-Autoskill framework transforms skill management into a unified lifecycle, consisting of five phases:

1. Creation: Generate skills on demand, with clear scope of application and interface contracts;
2. Memory: Store skill usage history (call context, execution results, feedback, etc.) to support experience accumulation;
3. Management: Classify and index, detect similarities, and enable intelligent retrieval to efficiently organize the skill library;
4. Evaluation: Use unit tests, runtime feedback, and quality metrics to ensure skill quality;
5. Optimization: Fix bugs, expand functions, and optimize efficiency based on evaluation and memory to form a closed-loop improvement. Skill-level memory is the core innovation, allowing skills to learn and grow from experience.

In the SkillsBench benchmark test, MUSE-Autoskill showed significant performance:

• Increased task success rate: Driven by evaluation mechanisms, intelligent skill selection, and memory to avoid failure patterns;
• Improved efficiency: Rapid skill reuse, optimized scheduling, and removal of inefficient skills;
• Enhanced reusability: Cross-task skill reuse reduces redundant creation;
• Cross-agent migration: Skills contain complete metadata and memory, enabling sharing between different agents and laying the foundation for a collaborative ecosystem.

Key technical implementation points include:

• Skill representation: Includes executable code + metadata (description, input/output specifications, dependencies, memory, etc.);
• Memory storage: Hierarchical strategy (hot data in memory, cold data on disk) + vectorized retrieval;
• Evaluation framework: Automated test generation and isolated execution;
• Optimization strategy: LLM analyzes memory data to propose targeted improvement suggestions.

Insights: Skills should be treated as long-term assets rather than tools; experience is transformed into improvements through memory; skill sharing promotes ecosystem building; balance quality and efficiency. Limitations: Memory bloat, privacy risks, skill conflicts, computational overhead. Future directions: Intelligent memory compression, skill market construction, version control, domain-specific applications (e.g., code generation, data analysis).

MUSE-Autoskill represents a significant advancement in skill management for AI agents, transforming skills from static tools into dynamic lifecycle assets. Skill-level memory allows skills to grow with experience, enabling true self-evolution. As agent deployment expands, effective skill management becomes critical, and MUSE provides a promising solution to this challenge, laying the foundation for the future agent ecosystem.