Agentic RAG is a retrieval-augmented generation system where an agent actively manages how retrieval happens during a workflow instead of following a fixed retrieve-then-generate pattern.
In a traditional RAG pipeline, the system retrieves context once and passes it to the model to generate a response. In agentic RAG, the model operates inside a broader control loop that can decide when retrieval is needed, what to search for, whether to rewrite queries, which tools or knowledge sources to use, how to combine evidence, and when additional retrieval steps are necessary.
This makes agentic RAG more flexible and capable for complex tasks like research, debugging, multi-document reasoning, and long-running workflows. But it also introduces more failure modes. Problems can come from poor retrieval quality, bad query reformulation, missing context, incorrect tool use, weak planning, grounding failures, or incorrect synthesis of retrieved information.
Because failures can happen across multiple steps, evaluating agentic RAG requires more than checking the final answer alone. Teams often use trace-level and trajectory-based evaluation to understand how the system searched, reasoned, retrieved evidence, and arrived at its final output.