RAG supplies the model with passages fetched from an external store at inference time, so answers can rest on material the model was never trained on. For agents, retrieval is a tool like any other, which means the agent chooses when and what to retrieve.