One morning I asked the browser agent to read a number off an ad dashboard, and all it sent back was the word "Sign in." In my everyday Chrome I was logged in and the dashboard opened as usual. The agent, though, could only see the front door.
I assumed the agent was broken and rewrote my prompt three times. It didn't help. The problem wasn't the prompt: the Chrome the agent drives is not the Chrome I use every day.
So this is less a bug than a design decision. If you're stuck at the same spot, here are four things to check, in the order I'd check them.
Which Chrome is the agent actually touching?
According to the official docs, Antigravity's browser agent opens and controls a local Chrome, using a completely separate Chrome profile. That isolation is meant to protect your personal data, so cookies and sign-ins from your normal browsing are not shared.
Being logged in on your Chrome therefore says nothing about the agent. Polishing the prompt won't change that.
The good news is that sign-ins made inside the agent's profile persist. Per the docs, your accounts stay signed in the next time the browser opens.
Check 1: Are you looking at the separate-profile window?
If Chrome is already running when the agent launches its browser, it shows up as a separate app with its own Dock icon (on macOS). If Chrome isn't running, it looks almost the same as your default profile.
Looks can mislead, so I check the launch arguments instead.
# macOS / Linux: list the profile directories Chrome is using
# Output varies by environment. Two distinct --user-data-dir values means two profiles.
ps -axo command | grep -i "[u]ser-data-dir" | grep -o -- "--user-data-dir=[^ ]*" | sort -uIf you only see one value, the agent's browser may not be running yet. Have the agent open it, then run the command again.
I do this because I'd rather replace "which Chrome is it?" with a concrete directory name before touching any settings. One fact first, guesses later.
Check 2: Has a human signed in once inside that profile?
This fixes it most of the time. In the window the agent opened, sign in to the service yourself. If there is two-factor authentication, I'd rather walk through it by hand than hand it to the agent.
After that, you can ask for things assuming you're logged in. I start with a deliberately tiny request:
Open the dashboard URL for the service and tell me only the account name or email shown at the top of the page.
If a sign-in page appears, stop and report "not logged in".It reads as little as possible and only tests the login state. Putting this step before any real extraction prevents the agent from reporting "nothing found" when it was simply logged out.
Check 3: Is the URL being blocked by access control?
The docs describe a two-layer model, a denylist and an allowlist, for which URLs the agent may visit. If pages after login won't open, you may be hitting this rather than a sign-in problem.
If the sign-in page loads but only the next domain fails, the service may bounce through a separate authentication domain, which large services often do. Check the official allowlist and denylist page and add only the domains you need. Not widening it too far is the main caution here.
Check 4: Did the profile location or a setting change?
The Browser section of the settings lets you change the profile location. If you moved it, you may be looking at a different profile from the one you signed in to earlier. If your login seems to have vanished, think back to whether you recently touched that setting.
The same section has a "Browser Tools" toggle. When it's off, the browser tools are disabled entirely, which is a problem before login even enters the picture.
| What you see | Suspect | First move |
|---|---|---|
| Only a sign-in page comes back | Not signed in within the separate profile | Sign in by hand in the agent's window |
| Signed in, but one page won't open | Allowlist / denylist | Find the blocked domain and allow only that |
| Worked before, suddenly logged out | Profile location changed | Check the location in the Browser section |
| Nothing browser-related runs | Browser Tools is off | Check the toggle |
Where I draw the line
Once I understood it was a separate profile, my real worry was whether to sign in there with my main account. Using an admin account is the quick way.
I use an account limited to viewing instead. That keeps the damage small if the agent misreads something. What I hand the agent is the range it can see, not my login. I try to hold that line even on rushed days.
A page that looks empty because it hasn't finished rendering is a different problem. I wrote about that in Why Antigravity's Browser Sub-Agent Reads SPAs as Empty Pages, and about day-to-day use in Two Weeks of Letting Antigravity's Browser Sub-Agent Handle My Weekly AdMob Mediation Review.
One thing to do today
Have the agent open its browser, sign in to your service in that window yourself, then ask it for the account name only, using the short request above.
If it answers, your login state is in order. If not, work through Check 3 onward; as an indie developer running dashboards at Dolice, that's still the order I follow myself.