From Static to Dynamic Skills: A Different Model for Agent Knowledge

Details

The post diagnoses a systemic problem in agent skill libraries: skills written as static markdown files freeze both procedure and facts at authorship time, with no dependency tracking, TTL, or invalidation protocol. When underlying data changes — a table gets deprecated, a join key shifts — the skill silently serves stale information.\n\nThe author (working on a multi-tenant AI data platform at Modus) argues that common remedies like owner fields, version schemes, and quarterly review cycles just add tidier metadata on top of knowledge that is still wrong. The underlying cause is architectural: a static skill is a cache with no invalidation mechanism.\n\nThe proposed alternative separates durable knowledge (intent, procedure, output contract, guardrails, scope definition) from ephemeral facts. Facts are resolved against a live context layer at the moment the agent asks, not at authorship time. The authored file becomes a build artifact with a lifetime of one call.\n\nThe post draws careful distinctions with RAG: unlike chunk-ranked retrieval with per-query metadata filters, the scope here is authored configuration compiled once per call into a shared filter clause, with re-validation on results. This prevents filter drift across multiple code paths and handles deleted-asset references more gracefully than similarity search.