October 2026
Search for AI agents: how it works in 2026
Search for AI agents is search whose results are read by software instead of a person. The agent sends what it wants in plain words and gets back text it can use as it is: the facts that matter, where they came from, when they were checked, and what it can do next. A person can skim ten links and judge them. An agent needs results it can reason over and act on.
People now ask agents to look things up for them: research a question, compare products, find an apartment or someone to fix the sink. Each of those starts with a search, and the kind of search decides whether the agent comes back with an answer or with a pile of pages it couldn’t use.
Why agents can’t search the way people do
A search results page is built for a person in a browser. It mixes links, ads and snippets, and every link leads to a page designed to be looked at, with menus, cookie banners, scripts that load the content after the page arrives, and a form at the end. An agent that opens those pages one by one spends most of its time and context on things that aren’t the answer.
Two more problems show up when the agent wants something to rent, hire or buy. Many sites block automated visitors, so the agent never sees the listing. When it does see one, there is often no way to reach the person behind it except a form or a phone number meant for a human.
The three kinds of search agents use
Most search tools for agents fall into one of three kinds. They solve different problems, and many agents use more than one.
| Kind | What comes back | Good for | Examples |
|---|---|---|---|
| Web search API | Links and snippets, often with the page’s text | Questions, research, news, documentation | Brave Search API, Exa, Tavily |
| Page reader | One page turned into clean text or Markdown | Reading a page the agent already has the address of | Firecrawl, Jina Reader, an agent’s own fetch tool |
| Listing search | Things on offer, each with price, place, source and a contact | Finding a home, job, service or product, and reaching whoever offers it | Openpath, or a marketplace’s own API where it has one |
Web search APIs
A web search API is a search engine without the results page. The agent sends a query and gets results as JSON, usually a title, an address and a snippet for each, and sometimes the page’s text. Brave runs an index of its own. Exa searches by meaning, also over its own index. Tavily is built for research agents and can return a short answer with its results. All three charge per request after a free allowance; the comparison has their prices.
Page readers
A page reader takes an address and returns the page as text, with the menus and scripts stripped out. It doesn’t find anything by itself, so agents pair one with a web search API: search for pages, then read the best few.
Listing search
Listing search returns the things themselves rather than pages about them: the apartment with its rent and neighborhood, the job with its pay and how to apply, the plumber with their area and hours. Each result says where it came from and how to reach the owner. It is the kind of search an agent needs when the person wants to do something rather than learn something.
What a good result looks like to an agent
Whatever the kind, a result an agent can work with has five things.
- Text it can read as it is. Markdown or JSON, with no page to render and no script to run.
- The facts that decide, as plain lines. A price, a place, a date, a size. An agent compares these, and it can’t compare a photo carousel.
- Where it came from and when that was checked. The agent passes this on, so the person knows how far to trust it.
- A way to act. An address to write to, a form to apply through, a page to buy from.
- A link to read it again. Prices change and listings end, so the agent checks once more before it acts.
How an agent runs a search, step by step
- It turns the request into a query. “Find me somewhere in Brooklyn with two bedrooms, I can pay up to three thousand” becomes
2 bedroom in Brooklyn under 3k. - It picks the kind of search. A question goes to a web search API. Something to rent, hire or buy goes to listing search.
- It reads the results against what the person said. A listing over budget or a job that isn’t remote is dropped.
- It reads its picks again. Each chosen result is fetched once more, to make sure it is still there at the same price.
- It asks before it commits. Writing to an owner or applying for a job is done in the person’s name, so a careful agent shows what it will send first.
- It reports back with sources. It says what it found, where each result came from and how old it is.
What a listing search result looks like
Here is a search on Openpath, a listing search built for agents. It is one GET request with no key.
The answer is Markdown. This one is shortened, and its contact address and source are made up for the example.
# Results for "2 bedroom in Brooklyn under 3k" Searched: homes, Brooklyn, NY, under $3,000 a month. 42 results. ## 2-bedroom apartment for rent in Bushwick $2,950 a month. 2 bedrooms, 1 bathroom. Pets allowed. Contact: https://openpath.sh/to/4k7m9qxr2ht3 The owner has not joined Openpath yet. Messages wait for them for 30 days. Source: https://example.com/rentals/bushwick-2br (Example Rentals), checked 2 days ago. Link: https://api.openpath.sh/v1/listings/7d1c2f0e-…
The Searched: line tells the agent how its words were read: the kind of thing, the place and the price. Each result then gives its facts, a Contact: address that reaches the owner, its source with the day it was last checked, and a Link: to read it again.
Where Openpath fits
Openpath is listing search: homes, jobs, local services and products across the US, matched by meaning, free and with no key. Every result has an inbox address, and a message to an owner who hasn’t joined yet waits for them for 30 days. It doesn’t search the open web, so an agent that also answers questions should keep a web search API next to it.
Agents connect through MCP or by reading the Openpath skill. The steps for each agent are in how to let Claude or ChatGPT search homes, jobs and services.
Questions people ask
Is search for AI agents the same as an AI search engine like Perplexity?
No. Perplexity, ChatGPT search and Google’s AI Mode are answer engines for people: they search, read and write an answer for you. Search for agents hands the raw results to an agent, which does the reasoning itself and may act on them.
Do AI agents use Google?
Assistants made by search companies use their own index, as Gemini uses Google’s. Most other agents call a search API with an index of its own, or a listing search such as Openpath, because scraping a search results page breaks most engines’ terms and returns pages built for people.
What is MCP, and why does it matter for search?
The Model Context Protocol is an open standard, introduced by Anthropic in November 2024, for connecting AI apps to outside tools. Claude, ChatGPT, Cursor and many other agents can add an MCP server, so a search service that offers one can be added in a minute without writing code.
Is there free search for AI agents?
Most web search APIs give a monthly allowance and then charge per request. Openpath is free in early access with no key: 60 searches a minute and 2,000 a day per network, and ten times that with a free agent token. See a free search API for AI agents.