Most agent skills today are static. Someone reads the truth once, writes it into a markdown file, and ships it. The procedure and facts get frozen together in the same paragraph. Then the table gets deprecated, and the skill doesn't know. A registry, owner field, version scheme, quarterly review, and linter only add tidier metadata to knowledge that is still wrong.

What we eventually admitted is that a static skill is a cache with no invalidation protocol. It caches a retrieval result behind a natural-language task description, with no dependency tracking, no TTL, and no way to notice that a source moved. That is why skills go stale, multiply, contradict each other, and bloat.

The alternative keeps intent, procedure, output contract, guardrails, and scope in an authored file. Every fact gets resolved against a live context layer at the moment the agent asks. The markdown the model reads still exists — it's just a build artifact with a lifetime of one call.

RAG with metadata filters stops one step short: the filter is a per-query argument assembled at the call site, and the unit of retrieval is a chunk ranked by similarity. In this model, a scope is authored configuration that compiles once per call into a single filter clause shared by every retrieval path and gets re-applied to results on the way out.

MCP resources move the fetch to call time, but selection stays with the client model, which must already know which resource to ask for from an authored enumeration. The staleness moves from the content into the index of what exists, and the model pays for its own exploration in round trips.

The deeper split is what gets ranked. Chunk similarity has no opinion about a column belonging to a table that belongs to a schema, so it cannot trade a table's full detail for three tables' identity lines when the budget gets tight. That trade is most of what the composer does.

This is a different model for agent knowledge: dynamic, live, and built around current context rather than frozen skill snapshots.