What does agentic RAG uniquely enable compared to naive single shot RAG?

**Agentic RAG** means retrieval is something the agent chooses to do, not a fixed step baked into the pipeline. In ordinary single shot RAG, the system takes your question, runs one search, stuffs the top results into the prompt, and generates an answer. Retrieval happens exactly once, before the model reasons at all. In agentic RAG, retrieval becomes a tool the agent can call inside its loop. The agent reads the question, decides whether it even needs to search, picks what to search for, looks at the results, and then decides whether to search again. It can do this many times in one task. This unlocks two things single shot RAG cannot do. The first is multi-hop retrieval, where the answer to one lookup feeds the next, like finding who wrote a paper, then looking up that author's other work. The second is iterative refinement, where a weak first search prompts the agent to rephrase and try again. The trade off is cost. Every extra retrieval and reasoning step is another model call, so agentic RAG is slower and more expensive than answering in one shot.

**Agentic RAG** reframes retrieval from a fixed pipeline stage into a tool call the agent issues inside its reasoning loop. The agent decides, at each step, whether to retrieve, what query to send, and whether the results are sufficient. Single shot RAG retrieves once, at query time, before any reasoning. That works when the relevant context is reachable from the original phrasing. It fails on questions whose answer depends on something you can only learn from a first retrieval. Agentic RAG uniquely enables three patterns. Multi-hop retrieval chains lookups, where the result of step one shapes the query for step two. Iterative refinement lets a weak first result trigger query reformulation and a retry. Adaptive retrieval lets the agent skip search entirely when it already knows the answer. The cost is real. Each retrieval adds a tool round trip plus a reasoning turn, so latency and token spend grow with the number of hops. You need a step budget and a stopping rule, or the agent can over retrieve. The design choice is therefore conditional. Use agentic RAG when questions are genuinely multi-step. Stay single-shot when one retrieval answers the question.