What is a computer-use agent?

You've probably talked someone through a website over the phone. Now click the blue button, top right. No, the other one. Some artificial intelligence (AI) tools can now do that job themselves. They look at the screen, move the pointer, click, and type, the way a person sitting at the computer would. A tool that works like that is called a computer-use agent.

Most assistants reach other software through tools: a clean, built-in connection that lets them ask a calendar or a search engine for something directly. But plenty of websites and older programs offer no connection like that. They were built for people, with buttons and forms. A computer-use agent gets around that by using the same screen you'd use. Anything a person can do with a mouse and keyboard, it can try.

Here's how it sees. The agent takes a screenshot, and the model looks at it as a picture. It finds the search box or the Next button, works out where it sits on the screen, and asks for a click at that spot. Then it takes another screenshot to see what changed. Click, look, type, look again. Anthropic, the company that makes the Claude models, described its version in October 2024 as looking at a screen, moving a cursor, clicking buttons, and typing text.¹³

That step-by-step looking makes it slow, often much slower than doing the task yourself. It can also misread a page, click the wrong item, or get stuck on a pop-up. Anthropic called its first release experimental, at times cumbersome and error-prone, and said scrolling and dragging were still hard for it. Agents have improved since. On one public test of everyday computer tasks, Anthropic's model scored about fifteen percent in 2024 and about sixty-one percent in 2025. That still leaves plenty of misses, and a busy page can still trip one up.¹⁵

An agent working inside a browser where you're signed in can reach whatever that browser can reach. Your email, your shopping accounts, saved passwords, a payment page. All of it. That's what makes it useful, and it's also why it needs limits. A web page can even carry hidden text meant to steer the agent somewhere you didn't ask it to go.³

Anthropic's guidance for developers lists the same precautions you'd want.

• Run it in a separate computer or browser, with as little access as possible.
• Keep sensitive things, like account logins, out of its reach.
• Have a person confirm anything with real consequences, such as paying, agreeing to terms, or sending.

If an agent you use offers a pause before payments or messages, keep it switched on.³

Anthropic introduced computer use for developers in October 2024. OpenAI launched ChatGPT agent in July 2025. It works in a browser on a virtual computer of its own and asks permission before steps like purchases. When a site needs a password, you type it in yourself, so it never passes through the model. Names and features change often. So before you hand one a task, ask yourself: what could it click, and where will it stop and check with me?¹²