Google rolled out a wave of fall updates to its Antigravity agentic environment: after the desktop app's May launch, version 2.30 on October 27, 2026 introduced a built-in browser, and version 2.31 on November 9 added strict security gates. Behind the headlines about AI driving a browser on its own, it's easy to miss the part that actually matters—the access rules deciding which buttons an algorithm is allowed to click and where the barrier drops.
Handing UI click-testing over to an AI sounds tempting, but an obvious question comes up right away: can you let a coding agent test live web interfaces without risking leaked passwords or accidental downloads? I cross-checked Julian Goldie's walkthrough against the Google Antigravity documentation to see how this sandbox actually works.
Why Blind Terminal Checks Aren't Enough for Live Layouts
Without a browser, a coding agent is like a chef who isn't allowed to taste their own food. Terminal linters and DOM checks catch typos in code, yet they're blind to real rendering: they won't notice a banner overlapping a button from a broken z-index, a mobile grid falling apart, or an animation freezing mid-frame.
Take a practical scenario: you ask Antigravity to build a React sign-up page with password validation. The agent compiles the bundle and spins up a dev server on localhost:5173. Previously, you had to open a tab yourself, type in a bad email, inspect the console, and paste screenshots back into chat. While Antigravity's /plan mode protects architecture at the start, Browser Use closes the blind spot at the finish line.
How the Browser Subagent Works in the Side Panel
When a task needs live verification on localhost:5173, the main agent spawns a Browser Subagent in the side panel (also triggered with /browser). It runs in a separate memory stream powered by the lightweight Gemini Flash model so it doesn't clutter the primary task's context, and pauses automatically if token limits are hit.
Here is what triggering a registration form check on localhost:5173 and configuring the gates looks like:
/browser Open http://localhost:5173, enter an invalid email in the sign-up form, and verify the red validation hint
# Browser Actuation Rules (Settings -> Browser)
Domain Allowlist: localhost:5173, localhost:3000, *.example.com
Unapproved Sites: Ask | File Downloads: Ask (Blocked: .exe, .dmg, .sh)
Once /browser fires, a cursor with a step badge moves across localhost:5173: the subagent types an invalid address, checks the red validation prompt, inspects console errors and network requests, and saves screenshots and video. It can juggle up to 3 tabs in parallel, and you can right-click any highlighted text to send it straight to the agent. Alongside this, November also brought Gemini Live voice mode, cloud agents, the Knowledge Work interface, and a preview of the /swarm command.
Security Gates, Port Allowlists, and Where the Boundaries Sit
You're probably thinking: "If the assistant clicks through forms on its own, what stops it from grabbing my Chrome cookies or downloading a malicious script?" That's where the security gates come in.
The Antigravity browser keeps an isolated profile and never touches your personal Chrome sessions. Login screens (Google, Microsoft, Okta, 2FA, passkeys, and CAPTCHAs) are off-limits to the agent: it pauses and waits for you to sign in manually and reply Signed In, continue. Download Interception catches window.open and download links and puts downloads on hold. Installers (.exe, .dmg, .sh) are always blocked, downloads cap at 20 files with manual approval, and system folders stay protected by Read-Only Storage. Even local addresses in Domain Allowlist are scoped by port: allowing localhost:3000 does not open localhost:8080.
The subagent runs in the Antigravity v2.30+ desktop app on Mac, Windows, and Linux for personal accounts (Free, Plus, Pro, Ultra), though it isn't available on Business or Enterprise plans or in cloud agents—and in the legacy IDE, /browser simply opens your regular Chrome. Arbitrary code execution on a page is blocked. You also need to be careful with the Turbo preset in Browser Actuation Rules: auto-approving navigation is handy on localhost:5173, but on third-party sites it opens the door to prompt injection hidden inside page markup (we covered access hygiene in our Antigravity plugin marketplace breakdown).
Before you start, configure your Deny list and leave external domains and file downloads set to Ask. Autonomous browser testing saves hours of manual UI QA only when a fast subagent is backed by strict port filters and a human gate on every password and file.

