Browser agent

How does a browser agent work?

OpenAI’s documentation describes a loop. The model receives a task, produces actions, the application executes them in a browser, and a new screenshot goes back to the model, which decides the next step. There are two ways to implement this. The model can write code in Python or JavaScript that drives the browser with libraries such as Playwright. It can also return structured actions such as click, type or scroll, which the application turns into browser input.

Anthropic describes its computer use tool in a similar way. Claude requests mouse and keyboard actions, the application runs them in a sandboxed environment and returns screenshots.

What does “headless” mean?

“Headless” means the browser runs without a visible window. Browser agents often run this way on a server. Some tools hand the agent a compact accessibility tree of the page instead of screenshots, for example agent-browser.

What does this mean for websites?

A browser agent loads pages in a real browser, so JavaScript is normally executed. Cloudflare lists “chat fetch bots and browser-use agents” in its agent category, next to search and training. How well an agent can be recognized in a log depends on the provider (assessment). Some label their requests, others look like an ordinary browser visit.

Both OpenAI and Anthropic advise treating page content as untrusted, since text on a page can contain instructions meant to steer the model.

Example

An agent is asked to book an appointment. It opens the home page, selects a location, fills in the form and submits it. If the form works only with a mouse drag or an unlabeled element, the agent may get stuck (Assumption).

FAQ

Is a browser agent a crawler? A browser agent works on a task for one person. A crawler runs independently across many pages.

Which products are browser agents? OpenAI’s ChatGPT agent counts as one. According to OpenAI’s launch post from July 2025, it uses a visual browser among its tools.

Last updated: October 8, 2026