How should agents route across thousands of skills?

Modern agents lean on skills — modular, reusable tool specifications (e.g. Anthropic's SKILL.md format) that bundle a capability with documentation for when and how to invoke it. As community skill libraries swell to thousands of entries, the obvious framing is a selection problem: given a query, which skill do I call? But real tasks rarely map to a single skill; they require composing several. That reframes routing as Compositional Skill Routing: decompose a complex query into atomic sub-tasks, retrieve the appropriate skill for each, and compose them into a dependency-aware executable plan (a DAG), rather than picking one tool off the shelf.

The reframing matters because it moves the hard part upstream. Selection accuracy is bounded by how well the query was carved into sub-tasks in the first place — a bad decomposition asks retrieval to find skills for the wrong steps. This is the skill-library instantiation of a principle the vault already holds: since Does separating planning from execution improve reasoning accuracy?, treating decomposition as its own stage (with its own failure mode and its own fixes) beats folding it into selection.

It also connects to how skill repositories should be built and maintained — Can a separate trained curator improve skill libraries better than frozen agents? — because composition quality depends on skills being scoped and documented well enough to be retrieved and chained. Compose-not-select is the demand side of the same skill-as-first-class-artifact shift.