Because attention is finite and instructions compete with each other, a system prompt holding every practice a team has ever written down performs worse than a short index that points to detail the agent can fetch. Progressive disclosure is the pattern that follows: keep a small always-loaded core, then split the rest into skills, documents, or tool definitions the model retrieves when the task calls for them.
It applies to tools as well as text. Deferred tool definitions, where an agent must search for a tool’s full schema before it can be called, let a harness offer many capabilities without spending context on all of them at once. The practical test for any line of standing guidance is whether it is needed on most turns; if it is not, it usually belongs one level down.