They Are Not Static: A Survey of Dynamic Agent Skills

Yubo Li

Keywords: LLM agents, agent skills, self-evolving agents, skill libraries, lifecycle management

TL;DR: Agents succeed not by accumulating more skills, but by lifecycle-managing, verifying, abstracting, and maintaining evolving skill libraries so they compound into reusable capability rather than polluted memory.

Abstract: LLM agents increasingly externalize procedural knowledge into reusable skills: code, natural-language procedures, SKILL.md packages, workflow graphs, or learned adapters. This shifts adaptation from prompt or weight updates to library updates that change what future policies can retrieve, compose, execute, and trust. We survey a 94-paper 2023--2026 dynamic-skill audit set and argue that dynamic-skill systems are best understood as lifecycle-managed, verified, evolving artifact stores. The paper makes three contributions. First, we extend the options-style 4-tuple skill formalism to a 7-tuple $\langle \mathrm{applic}, \mathrm{policy}, \mathrm{term}, \mathrm{iface}, \mathrm{edit}, \mathrm{verif}, \mathrm{lineage} \rangle$ and a library transition $\mathcal{L}_{t+1}=\mathrm{Apply}(\mathbf{u}_t(\mathrm{trig}_t,\mathrm{signal}_t),\mathcal{L}_t)$ described by a ten-operator vocabulary. Second, we organize the corpus into eight lifecycle stages and eight system families, separating where each family invests engineering budget from what failure mode it leaves behind. Third, we synthesize five evidence-graded patterns: admission gates matter, verifier quality is load-bearing in skill-aware RL, flat retrieval can drop at moderate library scale, maintenance appears load-bearing after growth, and write-time abstraction often beats read-time alone. We translate these patterns into a five-item reporting checklist for future dynamic-skill papers. The most consistent message across the literature is not that more skills help, but that what enters the library, under what gate, and with what abstraction is what determines whether a growing store becomes a reusable capability substrate or a polluted memory.

Presentation Mode: Undecided at this time.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 1