Survey Maps Lifecycle of Evolving LLM Agent Skill Libraries

Large language model agents are increasingly relying on external, reusable procedures, often termed "skills," which can manifest as code functions, natural-language instructions, or learned adapters. This comprehensive survey examines how these skill libraries evolve over time, conceptualizing dynamic skill systems as "lifecycle-managed, verified, evolving artifact stores." The research audited 124 papers from 2023-2026 to synthesize patterns in how agents acquire evidence, propose updates, verify candidates, organize for retrieval, repair stale entries, and govern sharing. The survey provides three key analytical tools: a six-sense taxonomy distinguishing various types of "skills," an eight-stage lifecycle architecture detailing design decisions from evidence acquisition to governance, and a lightweight skill-record schema with a ten-operator vocabulary for comparing library updates. Key findings highlight the repeated importance of admission and repair processes, the material impact of verifier quality on skill-aware reinforcement learning, and the potential degradation of flat retrieval as libraries grow. The paper concludes with concrete reporting standards and open problems for evaluating dynamic skills as continuously changing libraries rather than static collections.