Large language model agents increasingly rely on skills to broaden their capabilities for complex tasks. To achieve scalable skill use, Selective Skill Exposure (SSE) is required: provide relevant procedural content when needed while avoiding unnecessary exposure that can disrupt reasoning. Existing approaches each have drawbacks. Progressive disclosure gradually reveals metadata as the skill library grows, introducing context noise. Retrieval‑based methods limit catalog exposure but rely on a static query from the initial task, which may misalign with capability needs that emerge during reasoning and still inject irrelevant skills. We introduce SkillBrowser, a Meta‑Skill‑managed framework that enables an agent to detect its current capability gaps, formulate targeted queries, and load skills on demand. Experiments across six SRA‑Bench subsets and five LLMs show that SkillBrowser improves overall accuracy compared to static retrieval and LLM‑based selection, especially in multi‑step tool‑use scenarios, and mitigates accuracy drops on already‑solvable instances. Benefits vary with model and task, with under‑invocation of useful skills and added latency as key limitations.
These findings demonstrate that Meta‑Skill‑driven SSE can support scalable skill repositories, offering a viable path for large‑scale skill‑using agents.
Review