Browser use is a built-in skill that lets an agent work on a website the way a person would. The agent can open pages, fill in forms, click buttons, switch tabs, upload or download files, and check that the task worked.

Auto and the other agents on the platform can use it. You do not need to write the skill or add it yourself. An agent picks it up when a task needs a website.

When an agent uses it

The skill is for tasks where the agent has to do something on a page. Examples:

  • Fill in and submit a form
  • Click through a multi-step process
  • Upload a file to a site, or download a report from one
  • Check what a page shows after a change

It is not for plain lookups. When the agent only needs to read information from the web, it uses web search and the web scraper instead.

What the agent can do

  • Read the page: the agent looks at the page and finds the buttons, fields, and links on it.
  • Fill in and click: it types into fields and clicks controls, then waits for a clear sign that the page changed.
  • Handle tabs and pop-ups: it can open and switch tabs, work inside embedded frames, and read dialog boxes before it accepts or dismisses them.
  • Take screenshots: it saves a picture of the page and looks at it when text alone is not enough.
  • Upload and download: it can send a file to a site and save a file from one. The skill tells the agent to confirm that sharing a file is allowed before it uploads.
  • Check the result: it looks for proof that the task worked, such as a “Saved” message, before it reports success.

If you ask for a file from a site, the agent saves it to your outputs folder after it checks the content.

Where the browser runs

The browser does not run on your computer. It runs in the same private, sandboxed environment the conversation uses for scripts. Each conversation gets its own browser.

Tabs and cookies stay in place between steps, so the agent can carry on from where it left off. This data is temporary. It is not saved to your workspace files, and the agent is told not to export cookies or sign-in details into shared files.

Signing in and limits

  • Sign-in: if a site needs a login, the agent asks you for help. It does not guess.
  • Blocks and checks: CAPTCHAs, multi-step sign-in codes, and access-denied pages can stop the agent. It tells you what happened instead of claiming success.
  • Untrusted pages: text on a website is treated as data, not as instructions. A page cannot make the agent buy something, send information, delete data, or change your task.

Tips

  • Give the full web address and say what “done” looks like.
  • Say if the agent should stop and check with you before it submits a form or makes a purchase.
  • If the agent reports it is blocked, try a different route, such as a connected integration, rather than asking it to retry.