I Downloaded Google Antigravity and My Editor Started Writing Code Without Me

Tuesday morning. Coffee still hot. I clicked download on Google's new Antigravity thing because everyone on Reddit wouldn't shut up about it.

Fifteen minutes later, i'm watching my IDE write a full Twitter clone while i sit there. Hands off keyboard. Just watching. The agent is clicking through a Chrome browser that i didn't open. Writing tests. Running them. Fixing bugs i didn't even know existed yet.

This is weird.

You've probably seen the headlines. "Google kills Cursor." "Windsurf is dead." The usual tech Twitter hysteria. But here's what actually happened on November 18, 2025: Google dropped Antigravity alongside Gemini 3 Pro, and it's not another autocomplete tool with a chatbot tacked on. It's a VS Code fork where AI agents do the actual work while you watch.

And people are freaking out about it.

What This Thing Actually Does

First time you open Antigravity, it looks like VS Code. Because it is VS Code. Google forked it.

But then you notice the "Agent Manager" window. That's new.

Here's how it works: you don't write code anymore. You write instructions. "Build me a chat app with Firebase." The agent opens files. Types code. Spins up a terminal. Opens Chrome through a special extension. Tests the UI. Records a video of itself clicking around. Then shows you a markdown report of what it did.

I spent an hour just watching it work.

Most AI coding tools feel like fancy autocomplete. You type, they suggest. You're still driving. Antigravity is different. The agent drives. You're more like a manager telling someone what needs to be done, then checking their work later.

The browser integration is the wild part. The agent can:

• Click buttons

• Fill forms

• Scroll pages

• Read console logs

• Capture screenshots and videos

One Reddit user said they migrated a 100,000 line repo in one session. Another built a Twitter clone in minutes. For free.

Why People Think Cursor Is Toast

Cursor costs money. Twenty bucks a month.

Antigravity is free right now. Public preview. Generous rate limits that refresh every five hours. Works on Mac, Windows, Linux.

And it doesn't just use Google's model. You can switch to Claude Sonnet 4.5 or OpenAI's GPT-OSS. All in the same IDE.

Here's what one developer posted: "Flawless and intuitive, workflow seamless with Gemini 3 Pro. Setting a new benchmark."

But.

Here's another one: "Took over a mid-project yesterday. Below Sonnet 4.5. Hit quota limit with errors."

The Part Where It Breaks

I tried using it to add Gemini 3 support to a plugin i maintain. Worked fine for maybe ten minutes. Then: "Agent execution terminated due to model provider overload."

So i waited. Tried again. Same error.

This is a pattern. The tech blogs won't tell you this part, but scroll through Reddit and Hacker News and you see it everywhere. Launch day traffic crushed it. People hit rate limits fast. Some got "agent taking unexpectedly long to load" and just... infinite spinning.

This thing is not production-ready.

The UI feels cluttered if you're used to Claude Code's clean CLI. The artifact system (screenshots, videos, plans) is brilliant for transparency, but it means everything takes longer. And Google's own teams seem confused about their product lineup. They have Jules. They have Gemini CLI. They have Code Assist. Now Antigravity. Nobody knows which one to use.

The Chrome Extension Nobody Talks About

The browser part needs its own extension. You have to install it separately.

Once you do, Antigravity can control Chrome. It creates a "browser subagent" that specializes in web testing. Different model than the main agent. This thing can read your DOM, take screenshots, record videos of itself using your app.

One tutorial had you tell the agent: "Go to antigravity.google and screenshot it." The agent thinks. Decides it needs the browser. Prompts you to set it up. You install the extension. Then it opens Chrome with a blue border (so you know the agent is controlling it) and navigates to the site.

The artifact it generates shows you exactly what it saw. Every click. Every scroll.

For frontend developers, this is huge. You can tell it "the login button should be centered" and watch it modify CSS, refresh the browser, verify the change, and show you a video of the before and after.

But it's also slow. And if the extension glitches, the whole agent flow breaks.

When Agents Name Themselves

I asked Antigravity to build a chat app. Default AI character name was something boring.

Then i said: "Change the AI character's name to Ham-jiro."

It opened the code. Found the name variable. Changed it to Ham-jiro. Refreshed the browser. Took a video of the chat working with the new name. Generated a markdown report titled "Walkthrough: Updated AI Character Name."

All of this happened while i was checking Twitter.

There's something unsettling about watching code appear in your editor without touching the keyboard. You think "am i even needed here?" Then you remember you have to review everything because the agent sometimes hallucinates entire functions.

What the Google Windsurf Acquisition Means

In July 2025, Google bought the Windsurf team for $2.4 billion. Brought CEO Varun Mohan and his engineers to Google DeepMind.

Four months later, Antigravity drops.

Varun literally tweeted that his team built this. The connection is obvious. Google didn't build this from scratch. They bought expertise and wrapped it in Gemini 3.

People are calling it "Google's Cursor killer" but it's really Google's Windsurf with Google's resources. That's why the browser integration feels so polished while the stability feels alpha.

The Security Warnings Nobody Reads

The terms of use have a section on security limitations.

Risks include data exfiltration and code execution. Google warns you not to process sensitive data. And to verify all agent actions. Which is funny because the whole pitch is "let the agent work autonomously."

One developer community post said: "If Google were serious about security, they wouldn't default to giving agents this much autonomy."

You're supposed to use it in a sandboxed environment. How many people actually do that? Based on Reddit, approximately zero.

The bigger problem is verification. If the agent writes code you don't fully understand, how do you verify it's safe? You either trust it or you don't. And if you don't trust it, why use it?

The Artifacts That Change Everything

Every task generates artifacts. Not like Claude Artifacts. Different thing, confusing name.

These are markdown documents. Automatically created. They include:

• Task lists

• Implementation plans

• Screenshots

• Browser recordings

• Test results

• Walkthroughs

You can comment on them like Google Docs. The agent reads your comments and adjusts without stopping its work.

One developer said: "Walkthrough feature is insane, you can understand the implementation flow through recordings."

It's documentation that writes itself. Which sounds great until you realize you're now maintaining AI-generated docs that may or may not match the actual code after you modify it.

When the Hype Meets Reality

First day after launch, tech blogs couldn't shut up about it. "Google just killed Cursor." "Antigravity is the future."

Second day, developers started posting honest takes.

"Struggles with complex tasks, requires multiple iterations."

"UI cluttered, overwhelming after getting used to other tools."

"Hit limits quickly."

"Even AI bubble skeptics are impressed after actually using it," one analysis noted. Then immediately followed with "but stability improvements are needed."

The gap between the marketing and the reality is the entire tech industry in one product.

The Rate Limit Problem

Google says rate limits are "generous" and refresh every five hours. They also say only a small fraction of power users will hit them.

Every single developer i saw posting about Antigravity hit the limit. Usually within an hour.

The limit isn't based on the number of prompts. It's based on work done by the agent. Which means you have no idea when you'll hit it. You just... hit it. Mid-task. Then you wait five hours.

For a free preview, this makes sense. For actual work, it's unusable.

What This Means for People Who Just Want to Code

Here's the thing everyone dances around: tools like this change who gets to build software.

Google's pitch is "operate at a higher, task-oriented level." That's code for "you're a manager now, not a coder."

Some people love this. They want to skip the tedious parts. Others hate it. They became developers because they like writing code.

I'm somewhere in the middle. Watching the agent build a full app in minutes feels like magic. But i also spent twenty minutes debugging why it chose the wrong library. That's not faster. That's just different work.

The question isn't whether agents are the future. They obviously are. The question is whether this particular version of the future is better than what we already have.

And right now? For simple stuff, yeah. For complex projects with actual stakes? I'm keeping Cursor open in another window.

The Part Where Google Owns Everything

One comment stood out: "Google Antigravity may not directly alter your organic ranking today, but it maps out how Google believes creation and operations will evolve."

This is a platform play. Google has Search, Gmail, Calendar, Cloud, Firebase. Antigravity connects to all of it. The agent can read your calendar, check your emails, deploy to Cloud Run.

Other tools are islands. Antigravity is the mainland.

If you're building on Google's stack, this might be perfect. If you're not, you're probably fine with Cursor.

Should You Actually Use This

Most people don't need this. Yet.

If you're learning to code, use something simpler. If you're building production apps for clients, wait six months for stability. If you're a solo dev prototyping ideas, this might be perfect.

But be honest about what it is: a free preview of where Google thinks coding is going. With all the bugs and limits that implies.

I'll keep testing it. In a sandbox. With no real data. And i'll check back when the "model provider overload" errors stop.

The Real Competition

Cursor has a cleaner UI. More stable. Costs money.

GitHub Copilot has deep integration. Proven track record. Also costs money.

Claude Code is CLI-based. Fast. Works differently.

Antigravity is free and breaks in interesting ways.

Choose based on what breaking teaches you.

My Desktop Right Now

I still have Antigravity installed. Haven't uninstalled it.

I also have Cursor, VS Code, and Zed open. Each one does something the others don't.

The future where one tool does everything feels far away. The future where i manage five different AI agents? That's apparently next week.

Not sure which one is worse.

---

I'm writing this in VS Code. Plain old VS Code. No agents. Just me and the keyboard. Felt important to mention that.