Agentic RAG is one of the most active agent engineering directions in H1 2026. Unlike traditional RAG, Agentic RAG delegates retrieval decisions to the agent itself rather than a pre-defined fixed pipeline. Existing Agentic RAG patterns include: Introspective Retrieval — the agent first evaluates its own knowledge boundary, triggering retrieval only when uncertainty is high. Tool-assisted Retrieval — the agent uses multiple tools (search APIs, database queries, file system reads) via MCP tools instead of calling a vector database directly. Iterative Refinement Retrieval — the agent proposes more precise sub-questions or filtering conditions based on initial results, performing second or even third retrieval rounds until satisfactory results are obtained.
In the Zero to Codex course, Agentic RAG is one of the core use cases for agents calling MCP tools. The first practical MCP tool learners build is an 'intelligent retrieval tool' — accepting natural language queries, internally performing query rewriting, multi-path retrieval, and result ranking. Key engineering insight: after delegating RAG control to the agent, retrieval quality no longer depends on static pipeline tuning, but on the agent's strategy choices (when to retrieve, what to retrieve). The course recommends starting with 'single-step introspective retrieval' and gradually progressing to 'multi-step iterative refinement retrieval'.