Maxim Mironjuk
-
February 20, 2026
Classic RAG retrieves documents once and answers the question with the context it found. Agentic RAG goes a step further: Claude itself decides whether a first retrieval pass is enough, whether a follow-up query with refined search terms is needed, or whether multiple distinct knowledge sources must be combined. This article shows how to build such pipelines with clean tool definitions, where their limits lie, and how to debug multi-step chains when something goes wrong.