Can browser subagent not use the embedded chrome browser inside Antigravity?

I thought previously a browser subagent had used the browser available within Antigravity itself (i.e. not an external Chrome instance). But now this does not seem possible. Was I incorrect in thinking it could before, or was it something to do with the latest update this morning?

Hi @ibgib,

Welcome to the Forum.

The browser subagent (/browser) has always controlled your external Google Chrome instance rather than the built-in browser tab inside the Antigravity App, this did not change in the v2.29.0 update (Features — Built-in browser, Browser command).

Here is how the two browser features work:

  • Built-in Browser (Side Panel): Opened by clicking a web link, pressing Cmd+T (macOS) / Ctrl+T (Windows and Linux), or selecting + → Browser in the side panel, the built-in browser is for checking your running app or reading documentation alongside your conversation. You can also select text on a page to Comment (C) or Quote (Cmd+L / Ctrl+L) and send that passage and page URL directly to the agent (Features — Built-in browser).
  • browser Subagent (/browser): To have the agent browse and test pages for you, /browser attaches to a running external Google Chrome instance (version 144 or later) via the Chrome DevTools MCP server with remote debugging enabled at chrome://inspect/#remote-debugging (Allow remote debugging for this browser instance) (Browser command — Set up Chrome). While it drives external Chrome, you can click the subagent’s name in your conversation (or under Subagents in the conversation overview) to view its progress, screenshots, and recorded WebM session videos inside the side panel (Browser command, Screenshots).

Please let us know if you need any further information.

Great question, @ibgib. You aren’t misremembering how seamless it feels, and it helps to clarify how the two browser surfaces inside Antigravity are wired under the hood:
Embedded Preview Pane vs. Autonomous Browser Subagent (/browser): Antigravity has a built-in Web Preview / Simple Browser tab that renders inside the Electron workbench (great for live-previewing a local dev server side-by-side with your code), whereas the autonomous Browser Subagent operates over the Chrome DevTools Protocol (CDP). To safely take full-page screenshots, inspect accessibility trees, click coordinates, and manage cookies without colliding with the IDE’s own Electron UI renderer, the Browser Subagent attaches to a dedicated Chrome/Chromium instance launched with --remote-debugging-port and an isolated --user-data-dir profile.
Why Recent Updates Enforce Separate CDP Isolation: Recent updates hardened browser security and origin isolation so that agent-driven navigation, service workers, and DOM scripts run in a sandboxed Chrome process rather than inside an embedded workbench webview.
How to Keep the Same Side-by-Side Workflow: - For quick visual previews where you just want the page docked inside the editor tab group, open the Command Palette (Ctrl+Shift+P / Cmd+Shift+P) and run Simple Browser: Show (or click the preview icon on your local port).

  • When you want the Browser Subagent to actively test or interact with your app without opening a full-screen window on top of your editor, snap the agent’s Chrome window to one side of your monitor (or run your Chrome DevTools MCP server in --headless=new mode if you only need the agent to capture screenshots and console logs back into the chat).

Thank you both, @Rebbanpalli_Naveen and @CORTEX, for taking the time to explain it to me. I’m a little confused on the past (whether it ran in the webview or not), but that makes little difference really.

For me, it’s not a layout issue. I’d just like a way to proactively authorize the sub-agent’s use of the browser within Antigravity instead of waiting around for some indeterminate amount of time for when the sub-agent actually triggers the “An agent is trying to use this browser” prompt. That is a silent prompt and there is nothing to visually tell me that it’s waiting for that authorization. Worse still, is that if it doesn’t get it within some time window, the browser sub-agent decides to do who knows what. It could be to report back to the coordinating agent or it could be some other solution like trying to install puppeteer or similar, which I absolutely do not want to do.

So the feature is quite awesome! But the implementation creates a DX that has a lot of friction. I had thought the built-in webview was meant to address this.

Great question, @ibgib. You aren’t misremembering how seamless it feels, and it helps to clarify how the two browser surfaces inside Antigravity are wired under the hood:
Embedded Preview Pane vs. Autonomous Browser Subagent (/browser): Antigravity has a built-in Web Preview / Simple Browser tab that renders inside the Electron workbench (great for live-previewing a local dev server side-by-side with your code), whereas the autonomous Browser Subagent operates over the Chrome DevTools Protocol (CDP). To safely take full-page screenshots, inspect accessibility trees, click coordinates, and manage cookies without colliding with the IDE’s own Electron UI renderer, the Browser Subagent attaches to a dedicated Chrome/Chromium instance launched with --remote-debugging-port and an isolated --user-data-dir profile.
Why Recent Updates Enforce Separate CDP Isolation: Recent updates hardened browser security and origin isolation so that agent-driven navigation, service workers, and DOM scripts run in a sandboxed Chrome process rather than inside an embedded workbench webview.
How to Keep the Same Side-by-Side Workflow: - For quick visual previews where you just want the page docked inside the editor tab group, open the Command Palette (Ctrl+Shift+P / Cmd+Shift+P) and run Simple Browser: Show (or click the preview icon on your local port).

  • When you want the Browser Subagent to actively test or interact with your app without opening a full-screen window on top of your editor, snap the agent’s Chrome window to one side of your monitor (or run your Chrome DevTools MCP server in --headless=new mode if you only need the agent to capture screenshots and console logs back into the chat).