MMOOZZGG.com

How to build an AI agent

AI market research agent: search, summary and fact checking

5 min read6 sources

An AI market research agent searches open sources, reads what it finds and writes a summary with links: a market overview, a list of competitors, prices and industry news. It works by the ReAct approach of 2022, alternating reasoning and action, and you can build it in an open builder in a week. Still, check every number in its summary against the link.

On this page · 4 sections

Build it yourself in a week

How the agent searches and reads

A research agent works in a loop: it decides what to search for, searches, reads pages and decides whether it has enough. The ReAct approach describes this loop: the model alternates between reasoning and an action with a tool.

It needs two tools: search, which returns a list of links, and reading a page by its link. Large model vendors build web search in themselves: OpenAI added it to its Responses API on 11 March 2025, together with file search.

In n8n, an “HTTP Request” node connected to the agent as a tool reads a page. Its “Optimize Response” option with the HTML type and the “Return Only Content” setting keeps the text without markup, and “Truncate Response” cuts a page that is too long, so you do not pay for extra text.

Search is connected as a separate tool, through the API of a search engine for which you have a key. For the Russian market, it matters that the search finds Russian-language sources well.

The build in n8n is simple: an “AI Agent” node, a model and two tools. In the “System Message” you tell the agent which sources count as reliable and in what form to deliver the summary. The “Max Iterations” parameter limits the number of attempts, 10 by default, so the agent does not search forever.

For overviews across dozens of sources, people build a team of agents: a lead agent splits the question into parts, and helpers search in parallel. In Anthropic’s study of June 2025, such a system beat a single agent by 90.2% in an internal evaluation. You pay for this in tokens, the pieces of text by which a model is billed.

A summary with links

A summary without links is useless for a decision: the CEO does not know what to trust. So the agent’s instructions include a rule: every statement with a number comes with a link to the page where that number appears.

Anthropic set aside a separate agent for this in its system. After the search, it goes through the text and adds exact source links to the statements. In a simple agent, a second pass of the model over the finished summary does the same job.

Set the summary format in advance: the question, a short answer in 3–5 points, a table of numbers with date and source, and a list of what could not be found. The last item matters: gaps are visible at once, not hidden in smooth text.

The agent always writes the date of the source. A competitor’s price from two years ago looks the same in a summary as yesterday’s, and without a date you cannot tell them apart.

The instructions also list what the agent must not search for: forums without authors, retellings of retellings and pages without a date. A short list of bans saves time on checking.

For repeated overviews, such as weekly monitoring of competitor prices, save the format as a template. Then you can compare the summaries of different weeks line by line.

Fact checking

AI search gets links wrong more often than it seems. In March 2025 the Tow Center at Columbia University tested 8 AI search tools on 1,600 queries: together they answered more than 60% of the queries wrongly. For two services, more than half of the answers linked to invented or broken addresses.

A made-up fact presented with confidence is called a model hallucination. The second trap is the quality of sources: Anthropic says early versions of its agents picked pages made for search engines instead of authoritative sources, and it fixed this with rules in the instructions.

1. Open the link next to each number and find that number on the page.

2. Check the publication date and the author: a company, a study, a media outlet or a retelling of someone else’s text.

3. Check numbers that affect money against a second, independent source.

4. Mark in the summary what is checked and what is not, and do not pass on anything unchecked.

You can give part of the check to a second model: it compares a statement with the text behind the link. How to test an agent on a set of questions is covered on the page about agent testing.

Limits

The agent sees only open pages. Paid reports, closed databases and your managers’ talks with clients are out of its reach, so its conclusions about the market are incomplete without them.

Research costs money. According to Anthropic, agents use about 4 times more tokens than a normal chat, and teams of agents about 15 times more. A deep overview is worth it when a decision depends on the answer.

The agent does not know your context. Give it the company’s margins, plans and limits in the instructions, or it will deliver a general market overview instead of an answer to your question.

Search queries sent through a foreign model vendor go to that vendor. Do not send client names, internal figures or deal plans to such a search.

The agent does not replace an analyst where you need a call to a supplier, a talk with a client or a site visit. It cuts the rough part of the work: finding, reading and summarizing, while the conclusion and the decision stay with a person.

Sources and fact-checking

Checked by the editors on 16 October 2026.

  1. OpenAI, 11 March 2025 (openai.com)
  2. arXiv 2210.03629, 6 October 2022 (arxiv.org)
  3. Anthropic Engineering, 13 June 2025 (anthropic.com)
  4. Columbia Journalism Review, 6 March 2025 (cjr.org)
  5. n8n documentation: HTTP Request node, checked 16 October 2026 (docs.n8n.io)
  6. n8n documentation: AI Agent node, checked 16 October 2026 (docs.n8n.io)

Found a mistake? Tell us

Publications

Contact us