SkillSelect-Serve boosts small LLM agents with budget-aware skill composition

Researchers from multiple Chinese institutions introduced SkillSelect-Serve, a framework that reformulates how small LLM agents select and compose skills. Instead of returning a fixed list of top-k relevant skills, the system treats each skill as a structured Skill Service with attributes like functional description, dependencies, context cost, risk, and QoS metrics. A local Micro-Agent Requirement Planner converts natural-language tasks into structured service requirements, then a shared discovery backbone retrieves candidate services from a large registry. The framework performs dual-granularity utility modeling: skill-level marginal suitability estimation and bundle-level calibration to balance coverage, redundancy, cost, and risk.

Experiments using 35,353 skills and 586 task queries demonstrate consistent improvements over traditional fixed top-k retrieval baselines. SkillSelect-Serve achieves higher bundle recall and mean utility under the same budget, making it particularly valuable for deploying capable LLM agents on constrained hardware or with limited API costs. The approach is published on arXiv and could significantly enhance how small agents handle complex real-world tasks without requiring expensive large models or unlimited skill access.