A week inside Google Antigravity

I have been curious about Google Antigravity since it hit public preview back in November, and last week I finally stopped reading about it and ran it for five straight days on real client work. Not a toy. My main gig this spring has been a logistics company's dispatch board, a React front end over a FastAPI backend, wedged into a monorepo that crossed 400k lines a long time ago. If a tool can't survive that, I don't much care how it demos.

Antigravity is Google's agent-first IDE. It came out of the team and tech Google picked up from Windsurf, so underneath it's the familiar VS Code lineage with an agent layer run through the middle. It's powered by Google's Gemini 3-class models, though it can drive others. It runs on my M2 Ultra Mac Studio, on Windows, on Linux. It speaks MCP. And right now it's free in preview, which matters more than it sounds like it should.

the day I tried to use it like Cursor

My first day was a mess, and it was my fault. I opened Antigravity and immediately went hunting for the thing I live inside over in Cursor: fast, predictive autocomplete. I kept typing and waiting for the editor to finish my thought. It mostly didn't. I spent a few hours mildly annoyed, half-drafting a note that this was a weaker Cursor with a Google logo on it.

Then I gave up fighting it and just described a task instead. Rename a set of dispatch statuses across the backend, the API layer, and the three React components that read them. I hit go and went to make coffee.

That was the reversal. By the time I got back it had touched eleven files, run the test suite, and left me a short written plan of what it changed and why.

Antigravity is an agent with an editor inside it, not the reverse, and once I saw that the whole thing clicked.

what actually surprised me

The surprise wasn't code quality, which was fine, roughly on par with the frontier agents I already lean on. The surprise was where the agent went. It didn't stay in the editor. It opened a terminal and ran the migration. Then it opened a browser, loaded the dispatch board on localhost, clicked into the status filter, and checked that the rename actually held up in the running UI before it called the job done. I gave that browser loop its own write-up, because it's the part I keep turning over.

It's a genuinely different feeling to watch an agent work across the editor, the terminal, and a live browser in one motion. It behaves less like autocomplete and more like a junior who runs the thing before telling you it works.

There's a whole surface built around that, an Agent Manager for running several agents at once, and I gave the mission-control side its own piece, because that reshaped my week more than any single feature here.

where the preview shows

It's a preview and it acts like one. The browser step is slow, sometimes painfully so, thirty or forty seconds to do what I'd eyeball in three. It choked twice on our monorepo's build and I had to restart the agent both times. The artifacts it leaves behind, the plans and screenshots, are useful as evidence, but twice the screenshot showed one thing while the summary claimed another. You still read everything yourself.

None of that is damning for something this young. I'd just rather say it plainly than pretend the tool is finished.

the question a client asked first

Here's the part the demos skip. When I showed this to the client, their security lead didn't ask about code quality. She asked where the code goes, and which Google account it's tied to, and what happens to a private monorepo that an agent is driving through a browser on Google's infrastructure. Fair questions, all of them, and I didn't have crisp answers.

Then there's the money. It's free in preview as of this writing, and free is doing an enormous amount of work in every enthusiastic take you've read, mine very much included. An agent that drives a browser and runs a full test loop on every task is not cheap to run. That bill arrives eventually. Wire a team's daily workflow to today's price and you'll be meeting procurement down the line.

would I keep it open

For now, yes, off to the side. It hasn't replaced my main setup and I wouldn't hand it a production monorepo unsupervised. But it's the first tool that made me feel the shape of where this is heading, and that shape matters more than which editor wins the quarter. If you're weighing the field, I keep a running comparison of the coding tools I actually reach for.

Give it a real task, not a toy one, and watch where it goes.