Quick answer: An AI research agent is an LLM-driven system that answers a question by running its own research process: it breaks the question into sub-questions, searches the web, reads the sources it finds, takes notes, decides what it still doesn't know, and repeats until it can write a synthesized, cited answer. The defining trait is that the model controls the process. It decides what to search, what to read, and when to stop, instead of answering from one retrieval or from training data alone.
You've probably already used one. ChatGPT's deep research mode, Gemini Deep Research, Claude's research feature, and Perplexity's research mode are all research agents. They take minutes instead of seconds, and they come back with a report instead of a paragraph.
How It Differs From a Chatbot With Search
A chatbot with web search bolted on does one lookup, skims the snippets, and answers. That works for "who won the game last night." It falls apart for "compare the pricing models of the top five vector databases," because no single search result contains that answer.
A research agent treats the question as a project, not a lookup. The difference shows up in three behaviors:
| Behavior | Chatbot with search | Research agent |
|---|---|---|
| Queries issued | 1, maybe 2 | 5 to 30, reformulated as it learns |
| Reading depth | Snippets, sometimes one page | Full pages, cross-checked against each other |
| Stopping rule | Answers immediately | Stops when its sub-questions are covered |
| Output | A short answer | A structured report with citations |
The underlying loop is the same one we described in what is agentic search: search, read, assess, refine. A research agent is that loop scaled up, wrapped in planning at the front and synthesis at the back.
The Architecture: Five Stages
Every research agent I've seen or built, whether it's a commercial product or a weekend project, reduces to the same five stages.
1. Planning and decomposition
The agent first turns the user's question into a research plan. "Is llms.txt worth adopting?" becomes sub-questions: what is adoption like, does any crawler actually read it, what do implementations report, what are the counterarguments.
This stage matters more than it looks. A bad decomposition sends the agent down five variations of the same query. A good one covers the question from independent angles, which is what makes the final synthesis feel thorough instead of repetitive.
2. Search
For each sub-question, the agent issues web searches and reads the result listings: titles, URLs, snippets, dates. It's choosing what deserves a full read, so structured results matter. An agent reasoning over clean JSON picks better sources than one parsing a blob of HTML.
Good agents also reformulate. The first query teaches the agent the real terminology of the field, and queries two and three are usually the ones that find the good sources.
3. Fetching and reading
This is the stage most people underestimate. Snippets are not evidence. A snippet tells you a page might contain the answer; only the page itself contains the numbers, the caveats, and the publication date that make a claim usable.
So the agent fetches the pages it shortlisted and reads them in full, ideally as Markdown rather than raw HTML. A typical article is 3,000 tokens as Markdown and 40,000 as raw HTML, and that ratio decides whether your agent can afford to read fifteen sources or three. We covered the numbers in HTML to Markdown for LLMs.
This is also where the real world bites. Pages time out, sites block datacenter IPs, half the interesting content is rendered by JavaScript. Production agents route fetching through infrastructure that handles rendering and blocking, because an agent that can't read the page it chose is just a search engine with extra steps.
4. Note-taking and gap assessment
After each read, the agent extracts what's relevant into working notes: claims, numbers, quotes, and the URL each one came from. Then it asks the control-flow question that makes it an agent: what do I still not know?
Unanswered sub-questions trigger new searches. Contradictions between sources trigger tie-breaker searches. This assess-and-loop step is the feedback mechanism, and it's why research agents recover from bad first queries while single-shot RAG pipelines cannot.
Keeping notes instead of raw pages also solves the memory problem. Fifteen full articles won't fit in a context window alongside the reasoning. Compressed, source-attributed notes will.
5. Synthesis
Finally the agent writes the report from its notes: organized by theme, claims attributed to sources, disagreements surfaced rather than papered over. Citation discipline has to be enforced here, because a report without sources is indistinguishable from hallucination. We wrote about the techniques in how to make AI agents cite sources.
The synthesis step is where research agents earn their keep. Gathering ten sources is mechanical. Noticing that three of them repeat the same press release while a fourth contradicts it is the part that used to require a human.
Where the Quality Actually Comes From
Here's the thing I wish someone had told me before I built my first one: the model is rarely the bottleneck. Any current frontier model can run this loop. Quality differences between research agents come from three unglamorous places.
Source access. An agent that gets blocked by Cloudflare or can't render JavaScript reads a biased sample of the web: whatever happens to be easy to fetch. Its conclusions inherit that bias silently.
Token economics. Clean Markdown extraction is the difference between an agent that reads broadly and one that runs out of context after two sources. Depth of reading is a budget problem before it's an intelligence problem.
Stopping and verification rules. Uncapped agents loop forever or pad reports with junk sources. Good ones cap iterations, prefer diverse domains, and force a tie-breaker search when sources disagree.
All three are infrastructure and harness decisions, not model decisions. That's good news if you're building one, because it means the levers are in your code. A search endpoint that returns structured JSON and a fetch endpoint that returns clean Markdown, like the two provided by link.sc, cover the first two levers outright, and if you work in Claude or Cursor you can wire both in without code through the MCP server.
When You Need One (and When You Don't)
Research agents are the right tool for questions where the answer is scattered: competitive analysis, literature reviews, due diligence, "what changed in this space since last year." They're the wrong tool for lookups. Running a twenty-search loop to find a stock price is paying report prices for a fact, in both latency and tokens.
A useful rule: if a competent human would open more than three browser tabs to answer it, it's a research agent question. If one tab does it, a plain search tool call is faster and about a hundred times cheaper.
The Bottom Line
An AI research agent is a planner, a search loop, a reader, a note-taker, and a writer, glued together by a model that decides what to do next at every step. The gathering half is agentic search plus real page fetching; the synthesis half is note compression plus cited writing. The model matters less than you'd expect, and the quality of what the agent can search, fetch, and read matters more than almost anyone expects.
Give your research agent structured search and clean Markdown fetching in one API: create a free link.sc account with 500 requests a month included.