Tag Archives: browser use

Your Coding Agent Can Drive Your Browser. Copilot Just Made It the Default.

Last week I wrote about your AI coding assistant changing jobs — breaking out of one editor and following you into all of them. This week, the most mainstream one of them all learned to drive your browser. At Microsoft Ignite, GitHub put browser use into public preview across Copilot CLI, the Copilot app, VS Code, and JetBrains IDEs: your coding agent can now operate a real Chromium browser — click buttons, fill out forms, pull information from pages, and test and debug the web app you’re building, in the browser you’re already signed in to.

To be clear about what’s actually new here: agents driving browsers isn’t new. Anthropic shipped computer use in public beta in October 2024, OpenAI’s Operator followed in January 2025, and for well over a year the standard way to let a coding agent test your web app has been Microsoft’s own Playwright MCP server — one command (claude mcp add playwright npx @playwright/mcp@latest) and Claude Code or Cursor is clicking through your app. Anthropic even put computer use inside Claude Code itself back in March. What’s new this week is who’s shipping it and how: GitHub Copilot is the default assistant in a huge share of dev shops, and browser use is now built in across its whole surface — no MCP server to wire up, working in the Chrome you’re already signed in to, in an isolated tab group, off by default. The capability is two years old. The mainstreaming is this week.

That last phrase — “in the browser you’re already signed in to” — is both the feature and the warning label. Here’s what’s actually shipping, what it costs, and how to try it without handing an agent the keys to your logged-in life.

Copilot learns to click

GitHub’s changelog describes browser use as the next step in bringing computer use into everyday developer workflows. Paired with the new GitHub Copilot Chrome extension, Copilot can control the browser, debug web apps in real time, automate browser tests, and handle multi-step web tasks alongside the tools you already use.

The Copilot app’s Ignite release notes add the important guardrails: browser use runs in an isolated tab group, works in Chrome and Edge, and is off by default. In the Copilot app you have to turn on both Enable Computer Use and Enable Browser Use under Settings > Computer Use, and you can manage permission levels in the extension’s settings. Neowin’s write-up notes the same defaults: nothing happens until you explicitly enable it.

The practical requirements, per enjoy.dev’s walkthrough: Node.js 22 or later, Chrome 127 or later (or Edge with Chrome installed), and the GitHub Copilot extension from the Chrome Web Store. Then you flip the switch in your client — /plugin in Copilot CLI, Settings in VS Code and JetBrains — and start a new session.

Why this matters for a working developer: for teams that never wired up Playwright MCP or a computer-use setup, the loop just closed. Until now, your agent wrote the code and you became the test harness — alt-tab, refresh, click through the flow, paste the console error back. Now the same agent that wrote the form can submit the form, watch it fail, and read the page state itself. That’s not autocomplete anymore. That’s a junior dev with a browser — and now it’s in the assistant your company already pays for.

Microsoft’s own models join the picker — cheap

Ignite also put Microsoft’s homegrown coding models in your editor. DevOps.com reports that MAI Code 2 and MAI Code 2 Flash, from Microsoft AI, are available now in GitHub Copilot — Code 2 for deep debugging, multi-file changes, and repo-wide reasoning; Flash for explanations, targeted edits, and routine fixes. Digital Trends has the plan split: Code 2 on Copilot Pro, Pro+, Max, Business, and Enterprise; Flash on every plan including Free and Student.

The pitch is price. Microsoft’s claim is frontier-adjacent performance at a fraction of frontier cost — and “good enough, much cheaper” is a real strategy when your agent is burning tokens on browser test loops. Treat the benchmark claims as vendor claims until independent testing lands, but do check your model picker: under default settings these models can appear automatically, including in Copilot Fusion’s automatic routing. If your org keeps an approved-model list, this is the week to review it.

Claude Code changed a default you should know about

Meanwhile, Anthropic shipped Claude Code v2.1.328 on November 19, and the headline is a default. DevelopersIO’s breakdown confirms that auto mode now works for Claude Opus 4.6 and Sonnet 4.6 on Amazon Bedrock, Google Cloud’s Agent Platform, Microsoft Foundry, and the Claude apps gateway — and interactive sessions with no permission mode set now start in auto mode. Plan mode and Bash prompts on those providers offer auto mode too. If your team runs Claude Code through a cloud provider and your operating rules assume sessions start by asking permission, that assumption just expired. The off switch is disableAutoMode.

This is the same direction as browser use, from the other side: agents asking less often. That’s genuinely useful for unattended runs — and genuinely worth a deliberate decision rather than a surprise.

How I’d try it

  • Start in a throwaway profile. Browser use works where you’re signed in. Create a separate Chrome profile with only your dev/test accounts signed in, and let the agent drive that one. The isolated tab group helps, but a clean profile is the real bulkhead.
  • Point it at your own app first. “Open localhost, submit the signup form with bad data, tell me what breaks” is the perfect first task — high value, zero blast radius.
  • Keep permission levels tight. Leave the extension’s permission settings at their most restrictive until you’ve watched a few runs. Off-by-default is GitHub being careful; stay careful with them.
  • Check your Claude Code defaults. If you’re on Bedrock, Foundry, or Google’s Agent Platform with Opus/Sonnet 4.6, confirm your sessions start in the mode you intend — or set disableAutoMode.
  • Watch the model picker. Try MAI Code 2 Flash on routine edits and compare quality and cost against your current default before Fusion routes you there automatically.

A year ago, the coding assistant lived in a sidebar and suggested the next line. The browser-driving part isn’t new — but this week it stopped being a power-user setup and became a checkbox in the most mainstream dev tool there is, pointed at the browser you’re already signed in to. The tools are getting more capable and more autonomous at the same time. Your job hasn’t changed: know what you’ve actually turned on.