Generative Skill Composition tackles the growing bottleneck of picking the right agent skills
arXiv·high signal
As skill libraries grow, selecting the right composition of modular procedural-knowledge packages (sandbox setup, test running, multi-file refactors) becomes the central bottleneck for LLM agents. This paper frames skill selection as generative composition rather than flat retrieval or full-library exposure, directly relevant to anyone building Claude-style skill/subagent systems. It targets the exact scaling problem practitioners hit once a skill catalog exceeds a handful of entries.