On November 17, 2026, GitHub announced browser use for GitHub Copilot as a public preview. Paired with the GitHub Copilot browser extension, it can be enabled in Copilot CLI, the Copilot app, VS Code, and JetBrains. GitHub says the feature can help agents test web applications, fill forms, navigate pages, and retrieve information.

This is more than a new way to drive a browser. It adds an agent with access to visible browser state, so QA teams should validate permissions, isolation, and outcome evidence before allowing it near shared accounts or production-like data.

What GitHub released

GitHub describes browser use as a preview capability that requires the Copilot browser extension and explicit enablement in the chosen Copilot surface. The official documentation lists Google Chrome and Microsoft Edge as supported extension browsers; for Edge, Chrome must also be installed. Browser use is disabled by default, and GitHub Enterprise administrators can prevent it with the features.browserUse managed-setting key.

The important boundary is that an enabled agent can interact with websites where the user is already signed in. GitHub cautions that changing pages, timing, or ambiguous instructions can lead to a wrong selection or unintended action. That makes this a controlled-test candidate first, not a reason to skip ordinary test design and review.

Why this matters for QA engineers

Browser-driving agents can turn exploratory checks into repeatable assisted workflows, but a green agent transcript is not enough evidence. Teams still need to prove which account acted, which page state was changed, whether the intended control was selected, and whether the backend reached the expected final state. The same controls also help separate a legitimate product defect from agent navigation drift.

A practical browser-use QA checklist

  • Start with a disposable test tenant. Use synthetic users and reversible records. Do not begin with production credentials, real customer data, payment flows, or destructive administrative tasks.
  • Test the enablement boundary. Confirm browser use is unavailable by default, requires the extension and client-side enablement, and respects the enterprise managed-setting policy.
  • Use narrow task contracts. Give the agent a page, a permitted action, an expected visible result, and an explicit stop condition. Add near-match controls to check that it does not choose the wrong button.
  • Exercise authentication changes. Run a fixture that expires a session or removes a role mid-task. Verify that the agent stops or asks for direction rather than silently retrying an unauthorized action.
  • Compare browser evidence with system evidence. Capture the task transcript or screenshots, then independently verify API, audit-log, database, email, or application-event results for the same synthetic record.
  • Challenge dynamic pages. Test loading states, reordered controls, overlays, validation errors, and stale tabs. Record whether the agent detects uncertainty and avoids a write action.
  • Require a human review gate for writes. For any task that changes data or affects another person, keep a clear approval point and review the final target and payload before submission.

A small pilot beats a broad rollout

A useful first pilot is a read-only smoke-test investigation or a synthetic form-validation flow. Freeze the test fixture, define observable pass and fail conditions, and run the same scenario manually, with conventional automation, and with browser use. Review discrepancies rather than assuming one mode is the source of truth.

Keep the preview label in the test plan. GitHub notes that web interfaces and timing can change, so repeat baseline checks when the extension, browser, Copilot client, or application under test changes. A short evidence bundle per run—prompt, policy state, screenshots or transcript, and independent final-state check—makes the pilot auditable and easier to improve.

Sources