Researchers have built a framework that treats AI agent skill selection as a service recommendation problem, not a document retrieval one.
The paper, posted to arXiv, introduces SkillSelect-Serve — a system designed to help small language model agents pick and combine reusable skills from large libraries without blowing past compute or cost budgets. Instead of returning a fixed list of top matches, the framework converts a natural-language task into structured requirements, then selects and bundles skills by weighing functional fit against context cost, risk, and quality-of-service attributes. Tested against a library of 35,353 skills across 586 task queries, it outperformed fixed top-k retrieval baselines on bundle recall and mean utility at the same budget.
Most agent frameworks assume the model doing the picking is large enough to reason its way out of a bad skill selection. This work targets smaller models that can't afford that slack — meaning the selection layer has to do more of the cognitive heavy lifting. As LLM deployments move toward leaner, cheaper agents running at scale, infrastructure that constrains cost and risk at the skill level becomes load-bearing, not optional.
The approach borrows vocabulary from service-oriented architecture — dependencies, registries, composition — which is either a sign the field is maturing or a sign that enterprise software patterns are colonizing AI research. Probably both.