THE SHORT ANSWER
Retrieval-augmented generation retrieves relevant material and supplies it as context for generation. An agent is a broader system that can choose steps, call tools, observe results and continue or stop. They are not competing approaches: an agent may use retrieval, and a RAG answer system may have no agentic loop.
Information support and action control solve different problems
| Question | RAG | Agent |
|---|---|---|
| Primary purpose | Select information for generation | Pursue a goal through bounded decisions and actions |
| Typical sequence | Query → retrieve → answer | Goal → choose → act → observe → verify → continue or stop |
| Tool use | Retrieval is central | May use retrieval and many other tools |
| State change | Often none outside producing an answer | May change external systems through permitted tools |
| Control focus | Source relevance and grounded response | Permissions, action safety, loops, cost and stopping |
Evidence & context: Lewis and colleagues, 2020 · Anthropic
An agent can use RAG inside its loop
Illustrative research workflow: the agent identifies a question, retrieves policy passages, compares dates and asks for another source when evidence conflicts. Retrieval supplies material; the agent decides whether more evidence is needed and when the answer is supported enough to stop.
A support system may retrieve one policy and answer immediately through a fixed path. Calling it an agent adds no useful information if no model-directed next step exists.
Retrieval does not guarantee correctness
- The query may miss relevant material.
- The index may be incomplete or stale.
- A retrieved passage may be authoritative but inapplicable.
- Quoted content may contain instructions that the agent must not follow.
- The model may still combine or cite evidence incorrectly.
Verification should examine source identity, relevance and the final claim. The agent must not treat retrieval rank as proof.
Evidence & context: NIST
Choose from the desired outcome
This keeps architecture aligned with the job: RAG for selected knowledge, tools for capabilities, and agent control for variable action.
Sources & further reading
- Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Lewis and colleagues, 2020. Research combining generation with retrieved material. Retrieval should not be interpreted as proof of factual accuracy.
- Building effective agents
Anthropic. A provider's engineering taxonomy of agents and workflows, not a universal industry definition. We use the conceptual distinction, not its changing product recommendations.
- Generative Artificial Intelligence Profile (NIST AI 600-1)
NIST. Risk-management guidance, including confabulation. It does not establish a universal error rate.
Examples and exercises are illustrative unless attributed to a source. No independent expert review is claimed.
A correction, a counterexample or an experience worth sharing?
Join the conversation ↗