Copilot can use your browser now. Give it a rehearsal, not your whole workday

A browser task sounds simple until it reaches a button that actually does something. Opening a page is one kind of help. Publishing a change, sending a message or altering an account is another. The useful question about an AI browser assistant is not just whether it can click, but whether the task has a clear stopping point.
On November 17, GitHub announced browser use in public preview for Copilot CLI, the Copilot app, VS Code and JetBrains, paired with its Chrome extension. The announcement describes browser control, web-app debugging and browser tests. It also describes extension settings for permission levels. This is a new preview to evaluate, not evidence that every workflow will complete reliably. GitHub's announcement
Start with something you can inspect
Our suggested first task is deliberately ordinary: ask the assistant to inspect a test version of your own site and report a broken link. Use a disposable page or a staging copy containing invented data. You know what should happen, and a mistake does not become a public edit.
Write the boundary into the task: inspect this page, follow its navigation, report what fails, and do not submit forms or change settings. Then compare the report with what you can reproduce yourself. That gives you a useful baseline before you involve a real inbox, customer account or checkout.
A browser assistant that accurately notices a problem can save time even when a person handles the final fix. Autonomy is not the only useful outcome.
Connection is not completion
GitHub's setup instructions require the Chrome extension and enabling browser use in the chosen Copilot client. The extension should show Connected. In the CLI, GitHub also says to begin a new session after enabling the plugin. Follow the instructions for your client rather than assuming the menus are identical. Setup details
After connection, give the agent an observable finish line. For a layout check, that could mean naming the page, describing the overlap and providing the screen size at which it occurred. For a form rehearsal, it could mean stopping before submission and reporting the values it intends to send. These are editorial suggestions, not guarantees about the preview's enforcement.
Decide what must remain human
Before the first consequential task, write down the actions you want to approve yourself. Publishing, payments and account changes are sensible candidates. Keep the test browser free of unrelated sensitive tabs where practical, and review the extension permissions rather than treating a connected indicator as a privacy promise.
The payoff is a workflow you can trust in small pieces: inspect, propose, approve, verify. Our agent failure-check guide gives you more ways to think about that boundary. Today's companion briefing explains how repeatable tests can help judge custom AI behavior.
This is source-checked editorial guidance, not a hands-on review of the browser-use preview. Availability and controls may change during preview.