Classic retrieval-augmented generation runs one query, pastes the results into the prompt, and answers. That works when the user's question maps cleanly onto one lookup. It fails on questions that need two facts from two places, or that use vocabulary the corpus does not share.
Agentic RAG turns retrieval into a loop the model controls. It rewrites the question into search terms, runs a query, reads what came back, and decides whether that is enough. If a piece is missing it issues another query, possibly against a different source, before writing an answer.
Concretely, "which of our enterprise customers renewed after a P1 incident last quarter" becomes several searches: the incident log, the renewal records, then a join the model performs itself. No single embedding lookup would have returned that.
The costs are real. Each loop adds latency and tokens, and a model that searches badly can loop until it hits a limit. Production systems cap iterations, log every query for review, and fall back to single-shot retrieval for simple questions.