Retrieval as a tool the agent chooses to use
A common but wasteful pattern is stuffing retrieved context into every single prompt, regardless of whether the question actually needs it. In an agent, there's a better option: treat knowledge retrieval as just another tool in the agent's toolbox, callable the same way as any other function. The agent sees a lookup_documents-style tool alongside its other tools, and it decides, based on the request in front of it, whether it actually needs to call it. A question the model can already answer confidently from its own training never triggers a retrieval call at all, saving tokens, time, and the risk of introducing distracting or contradictory context nobody asked for.
This works because it reuses the exact same request-then-execute-then-feed-back mechanism as any other tool call. The model outputs a structured request for retrieval, your code runs the actual search against your knowledge base, and the results come back into the conversation the same way a database query's results would. Nothing about retrieval is special-cased here, it's simply one more capability the agent's loop already knows how to invoke, decide it needs, and incorporate before continuing.