Dear Google DeepMind Engineering Team & Developer Community,
I am writing this critical feedback regarding what used to be one of the most advanced and game-changing capabilities in Antigravity: the integrated Browser & Live DOM Recording system, as officially documented here:
Browser Overview | Google Antigravity Docs
Regrettably, in the latest update, the sophisticated workflow described in that documentation no longer functions as intended and has suffered a severe architectural regression.
1. The Groundbreaking Experience We Previously Had (Based on Docs & System History):
In earlier versions, Antigravity’s browser integration was truly autonomous and state-of-the-art:
- Integrated Browser & Active Tab Management (
execute_url): Antigravity previously utilized a dedicated standalone browser engine (with anantigravity-browser-profiledirectory). When launching the browser, both the developer and the AI had transparent visibility into all active tabs, could cycle between them effortlessly, and could execute URLs automatically without complex setup. - Live Visual Recording & Automated Screenshots (
browser_recordings): The system actively captured high-frequency visual frames (.webp/.pngstreams and action metadata inbrowser_recordings/). The AI genuinely “saw” UI/UX layouts, navigation movements, and visual styling defects on screen just like a human engineer. - Autonomous End-to-End Bug Fixing (Closed-Loop Hotfixing): The AI could spot an aesthetic or layout flaw on screen (e.g., overlapping text, broken flexbox, styling misalignments), immediately trace and modify the source code in the IDE workspace, and hot-reload the page to confirm the fix in one continuous, seamless motion.
2. What Broke in the Latest Release (After Migrating to chrome-devtools-mcp):
Migrating the browser subsystem to an external Chrome DevTools MCP setup has severely degraded the entire workflow:
- Standalone Browser Lifecycle Is Stripped: The built-in
execute_urltool andantigravity-browser-profilehave been abandoned. The agent is now prohibited from launching browser instances independently. Developers are forced into manual rituals: launching Chrome externally, openingchrome://inspect/#remote-debugging, and manually enabling remote debugging on port 9222 (DevToolsActivePort ENOENT). - The Agent Has Been Blinded to UI/UX: The live visual recording stream in
browser_recordings/and active tab visibility have vanished. The agent is now fed a static Accessibility Tree dump capped at just 8 KB. It can no longer meaningfully assess visual polish, responsiveness, or frontend layout bugs visually. - Severed Visual-to-Code Pipeline & Sandbox Violations: The browser is isolated into separate subagents plagued by strict timeouts (Navigation timeout 10,000 ms) and workspace sandbox path restrictions when attempting to save viewport screenshots (
Error: Access denied: path is not within workspace roots).
3. Our Request to the Product & Engineering Team:
The official documentation at https://antigravity.google/docs/ide/browser/ represents a brilliant vision for agentic pair-programming. Please restore the integrated, autonomous browser environment with active tab visibility (execute_url), native visual frame recording (browser_recordings), and tight in-editor hotfixing integration. Do not allow what was once Antigravity’s most cutting-edge flagship feature to regress into a rigid, blind text-dump utility.
Thank you for your hard work and continuous iteration. We sincerely hope to see this essential workflow restored in upcoming releases.
