Research Agent — AI Glossary

Reference details

Related terms

• Web Agent
• Agentic RAG
• Tool-Using Agent

How it works

1. A goal is turned into an initial set of searches, and the agent — not the operator — decides what to look for.
2. Retrieved pages are read and assessed, so a low-quality source can be discarded rather than cited.
3. Gaps are identified against the goal, and those gaps generate the next round of queries, which is what makes the process iterative rather than a single retrieval.
4. Findings accumulate in a working store with their sources, because a report asserting facts without them is not checkable.
5. The loop ends on the agent's own judgement of sufficiency, so it needs externally imposed caps on time, steps and spend.

Example: Why the cost is an order of magnitude, not a multiple

A RAG answer is one retrieval and one generation — two calls. A research run issuing 15 searches, reading 40 pages at roughly 2,000 tokens each, and synthesising, processes about 80,000 tokens of source material plus the reasoning over it. That is tens of calls against two, and minutes against seconds. The output is a different artefact — a sourced report rather than an answer — so the comparison is not like for like, but a budget built from RAG assumptions will be wrong by roughly an order of magnitude.

Commonly confused with

Not RAG. A RAG pipeline retrieves from a corpus you assembled and controls what the model can see; a research agent decides what to look for, chooses sources at runtime, and reads whatever it finds. That is why it works on questions you did not anticipate, and why the untrusted-content problem lands squarely on it — the pages it reads can contain instructions.

When to use it

Reach for it when:

• The question needs synthesis across sources nobody has assembled into a corpus.
• Coverage matters more than latency, and minutes of runtime are acceptable.
• The deliverable is a sourced report rather than a single answer.

Reach for something else when:

• The answer exists in a corpus you control, where RAG is far cheaper and more predictable.
• Interactive latency is required.
• No caps on steps, time or spend, leaving the agent to decide its own budget.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Research Agent — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path