[reggz]

In practice I’ve found agentic search does improve accuracy on harder queries, but the cost climbs pretty quickly once you move away from a single pass. More steps, more chances to call the model again, more room for retries. It compounds faster than you expect.

Plain RAG is easier to keep predictable if your retrieval is tight. But once chunk count grows and you start pushing larger contexts, quality drops and you end up compensating somewhere else.

What’s worked best for me is a hybrid. Cheap default path first, then only route ambiguous or high-value queries into the heavier agent flow.

Also I think cost per successful task is more meaningful than cost per query. The expensive path only makes sense if it actually reduces retries or downstream fixes.