SkillSelect-Serve Optimizes LLM Agent Skill Selection with Budget Control

Reusable skill libraries are becoming essential for large language model (LLM) agents, yet current skill selection methods often treat skills as simple documents, returning fixed lists without considering deeper operational aspects. This research presents SkillSelect-Serve, a novel framework that reframes agent skill selection as a sophisticated problem of skill service recommendation and composition. SkillSelect-Serve represents raw skills as structured "Skill Services," complete with functional descriptions, interdependencies, context costs, risks, and Quality-of-Service (QoS) attributes. A local Micro-Agent Requirement Planner translates natural language tasks into structured service requirements. A shared discovery backbone then identifies candidate services from a vast registry. The framework employs dual-granularity utility modeling, estimating marginal suitability at the skill level and calibrating bundles for coverage, redundancy, cost, and risk trade-offs. Extensive experiments involving over 35,000 skills and nearly 600 task queries demonstrate that SkillSelect-Serve consistently improves bundle recall and overall utility compared to fixed top-k retrieval baselines, especially under budget constraints.