RAG is the default story for grounding a model in your data: embed everything, retrieve top-k, stuff it into context, generate. It is a fine baseline. It is also where a huge number of prototypes plateau and get blamed on the model.
The failure mode is almost always the same. The user asks a question whose answer lives across multiple documents, or requires a sub-question to be answered first, or depends on a piece of information whose phrasing does not match the query. Single-shot retrieval cannot do any of that.
Agentic search is the alternative. The model treats search as a tool, reasons about what it found, decides whether to refine the query, follows up on a promising thread, and stops when it has actually answered the question. It is slower and more expensive per query. It is also dramatically better at the questions that matter.
The choice is not philosophical. Fast factual lookups want RAG. Investigative or multi-hop questions want agentic search. Most real products want both, with a router that picks. The mistake is shipping one and pretending the other class of question does not exist.