An AI agent that browses websites: what it can do on a site, and where it stops
Some sites have no export and no connection, only a search box. How a browsing agent works through them, when a simpler block is better, and the lines it never crosses.
Part of the guide AI employees
An AI agent that browses websites opens a real browser, does a task on a site the way a person would (types in the search box, picks a filter, opens a result, reads it) and brings back the fields you asked for. In xMap Flows this is the Browser block: you give it a start page and a task in plain words, and xMap AI drives the browser one step at a time. It is meant for sites with no connection or download: public registers, council portals, directories and suppliers' catalogues.
Browsing is powerful and slow, so the useful questions are when to use it, how to phrase a task it can finish, and what it will refuse to do. This article answers those for teams who keep copying the same information out of the same website by hand.
When browsing is the right tool
Pick the lightest block that does the job:
| You need | Use |
|---|---|
| Businesses, their ratings, phones and positions | Nearby places or Map business scan |
| To know when a page's text changes, or a list on it grows | Web page watch |
| Stories from news feeds or a news page | News |
| To search a site, filter, click through and read results | Browser |
Browsing is for tasks that need clicks and searches, not just reading one page.
Web page watch reads a page and compares it with last time, which is quicker and cheaper than browsing. The Browser earns its place when the answer sits behind a search form or two clicks deep.
How a task runs
The block takes a Start page (a public page, such as a portal's search screen), a Task written as you would tell a colleague, and the fields to Bring back, up to 12, each becoming its own field. Run it either once a run, or once for each incoming item, with {{field}} putting each item's values into the task.
Each click, search or page read is a step, up to the Most steps you set. A task has about two and a half minutes. The browser keeps to the start page's site and any sites the task names, and the AI is told today's date, so a task such as applications received in the last 7 days is read against the real calendar. A task that finds nothing fails the step and says what it saw, which is how you notice a site that changed its layout.
When you run the flow from the editor you can watch: a read-only window over the canvas shows the browser live with the step it is on. Scheduled runs have no window; their result says what the browser did in browser_summary.
The lines it does not cross
- It never types a password, a card number or a bank or identity number, and never presses buy, pay, order, book, sign up, agree or delete.
- By itself it only reads and searches. It sends a form, such as a quote request, only when you set Forms to send the forms your task asks for, and never in a test run.
- It signs in to a site only as you: you sign in yourself in a live window, and xMap never sees or keeps your password. Only flows you own use that sign-in.
- It leaves alone a site whose robots.txt asks robots not to visit, and treats words on pages as data, so a page cannot tell it to do something else.
- It keeps no recording of the pages it opened.
A worked example (made up)
A shopfitting company in Manchester wins work when restaurants and cafes apply for planning permission. The council's planning portal has a search form and no export. Every Monday, someone used to search it by hand. Now an AI employee does:
- Schedule: every Monday at 07:00.
- Browser, once a run. Start page: the portal's search page. Task: Search planning applications received in the last 7 days with the words restaurant or cafe, open each and read its details. Bring back: address, reference, proposal, status, date received. Most steps: 20.
- Only new, so an application seen last week is not sent again.
- xMap LLM scores each from 0 to 100 as a shopfitting lead from its proposal text.
- Email: one email to the sales team with the list, best scores first.
Monday morning's inbox holds a short list: each new application with its address, what is proposed and a one-line reason for its score. If the council redesigns its portal, the step fails with what the browser saw, and the team knows to adjust the task.
Writing a task it can finish
- Name the action and the end point: search for X, open each result, read Y.
- Give limits: the first 20 results or the last 7 days.
- Ask for fields the page really shows. A field it cannot find comes back empty.
- Keep one site per task. Several sites belong in several Browser blocks or one task per incoming item.
- Raise Most steps only when the step log shows the task running out of steps.
Mistakes to avoid
- Browsing a page that only needs watching: Web page watch does it for less.
- Using it on sites whose terms forbid automated visits. The bot-check option exists, but use it only where the site allows it.
- Expecting a free test: a test run uses the real browser and its steps are charged; only the estimate shown while you edit is free.
- Running it for each item on a long list. It handles a handful of items a run; filter first.
Only the steps of tasks that worked are charged, and a run never takes more steps than its credits can pay. For the other kinds of AI employees and how they fit together, see the AI employees guide; for a ready-made browsing employee, try what the best-rated cafes near you say on their own websites.
Questions
Can the browsing agent log in to a site?
Only as the flow's owner. Switch on Use my sign-in to the site, sign in yourself in the live window, and press Done. xMap never sees your password.
Will it fill in contact forms?
Only if you set Forms to send the forms your task asks for, and only with details written in the task. It never sends one in a test run.
How many pages can it read in one run?
Each task has up to 25 steps and about two and a half minutes. For each incoming item, it handles up to five items a run.
What happens when the site changes?
The task finds nothing, the step fails and says what it saw, so you can update the task or the start page.
Templates that do this
Read next
More on ai employees and automation · Every article · Templates