An Honest Review of Google Antigravity
Source: Dev.to
If you’ve been paying $20 / month for Cursor to get the best AI‑coding experience, you might want to take a look at Google’s latest offering: Antigravity. While it’s still in preview and completely free, it packs the newest Gemini models, full‑browser orchestration, and a brand‑new way to manage code—no credit card required.
TL;DR
- What it is: A VS Code‑based fork built by Google, re‑branded and extended with AI‑first features.
- Key concept: Three Surfaces – Editor, Browser, and Agent Manager.
- Strengths: Powerful Gemini 3/2.5 models, autonomous agents that can generate code and assets, integrated Chrome testing.
- Weaknesses: Rough prototype‑level bugs (UI glitches, broken extensions, flaky agent control).
1. The Backstory
For the past year the “king” of AI‑assisted coding has been Cursor, a $20/mo subscription. Google has now entered the ring with Antigravity, a VS Code fork that is 100 % free during preview.
“When Google releases a developer tool it’s either incredible or it dies in six months.”
I’ve been daily‑driving Antigravity for a week, digging into the code and the workflow. The result? It’s both one of the most promising pieces of software I’ve seen this year and the single most frustrating editor I’ve ever used.
2. Under the Hood
- VS Code base: Open‑source VS Code OSS is still there; you can see it in the About dialog.
- Re‑branding:
- “VS Code Settings” → Editor Settings
- Microsoft branding stripped out
- File‑search references a mysterious “Cascade”
Cascade is the name of the AI agent in the Windsurf editor. This suggests Google may have acquired (or forked) some Windsurf tech to accelerate Antigravity’s launch.
3. The Three Surfaces
Google’s thesis: an IDE should be an operating system for agents, not just a text editor.
| Surface | What it does |
|---|---|
| Editor | Classic code‑editing pane where you type. |
| Browser | A fully‑controlled Chrome instance embedded in the IDE. |
| Agent Manager | A separate window that acts as an Inbox for development tasks. It shows which agents are thinking, waiting for approval, or have failed. |
Agent Manager in Action
- Acts like Mission Control for your AI assistants.
- Supports multiple projects side‑by‑side (ideal for the “ADHD developer” who jumps between tasks).
- Shortcut ⌘ + E focuses the Editor on the selected project.
When it works, the workflow feels futuristic; when it glitches, it’s a reminder that we’re still in prototype territory.
4. Demo: Building Insaniquarium with Gemini 3
Prompt: “Create a clone of the old Insaniquarium game (fish eat food, drop coins).”
| Model | Result |
|---|---|
| Codex‑High (or similar) | Took an hour, burned ~3.5 M tokens, produced broken physics and 2‑D sprites floating in a void. |
| Gemini 3 (Antigravity) | One‑shot a working Phaser game and generated the required assets (fish, food, background). |
How the Agent Handled Assets
- Recognized it needed sprites.
- Paused the code generation, invoked its internal image‑generation model.
- Injected the generated PNGs into the project.
First attempt: sprites had white backgrounds (no transparency).
Fix: I told the Agent “Fix the transparency.” It regenerated the assets and updated the code automatically.
Agent‑First distinction: The AI isn’t just a chatbot; it’s a worker with access to tools (image generator, file system, browser) that you can direct like a junior developer who also does graphic design.
5. Browser Integration
Because Google builds both Chrome and Antigravity, the two talk seamlessly.
- Browser Agent: Spin up a Chrome window that runs a test suite.
- Live demo: “Go to
localhostand test the game.”- A blue Agent Control border appears.
- A red dot (the AI’s cursor) clicks, drops food, and verifies physics.
- Artifacts:
- Video recording of the test run.
- Screenshots of any errors (available in the Walkthrough tab).
6. The Bugs (and Why They Matter)
| Area | Symptom |
|---|---|
| General UI | Buttons stop responding; sidebar icons disappear until you click them repeatedly. |
| Extensions | The popular Svelte extension crashes the entire editor. |
| Agent Controls | “Controls Disabled” warnings appear even when the agent is active. |
| Port Management | With many projects open, the Browser Agent sometimes connects to the wrong localhost port. |
| Review Code Button | Clicking “Review Changes” instantly closes the dialog—feels like the editor is trolling you. |
These issues make Antigravity feel like a prototype rather than a polished product. The experience can swing from “mind‑blowing” to “exasperating” in a single session.
7. Final Thoughts
- Potential: The Agent‑First workflow, Gemini 3’s code‑and‑asset generation, and tight Chrome integration are game‑changing if they mature.
- Current state: The UI is flaky, extensions are broken, and the Agent Manager still has rough edges.
If you’re an early‑adopter who enjoys tinkering with bleeding‑edge tools, Antigravity is worth a deep dive—just be prepared for frequent crashes and a lot of manual “reset” work.
Stay tuned for updates as Google iterates on this ambitious IDE.
Google Antigravity – A Fascinating, Frustrating, Yet Futuristic Mess
“Wait when the Agent tells you it’s ready.”
Missing Features
- Arrow keys don’t work in the file explorer (you have to click) or in the code editor—an obvious usability issue.
Power‑User Grievances
- No Git Worktrees – Cursor has this, and it’s amazing for switching branches instantly. Antigravity doesn’t support it yet, which is ironic because the whole point of the Agent Manager is multitasking.
Efficiency Concerns
- The app feels heavy.
- Significant battery drain on a MacBook Pro while running.
- A glowing effect around the UI when the Agent is thinking looks cool but causes input lag; typing while the Agent generates feels noticeably delayed.
Why? You’re running a local Chrome instance, a heavy Electron app, and constant streaming connections to Gemini. It eats resources for breakfast.
So, Where Does That Leave Us?
Google Antigravity is a fascinating, frustrating, yet futuristic mess.
What Works
- The Agent Manager workflow—an “Inbox” for your code—is genuinely a great idea. I hope every other editor copies that.
- The Gemini 3 model is shockingly good at one‑shotting complex tasks and generating assets.
- Price? You can’t beat free. Google says the free tier limits reset every five hours. I hit those limits a few times, but for most people it’s plenty. You get access to state‑of‑the‑art models without a subscription.
What Doesn’t
- I wouldn’t uninstall Cursor for this—not today.
- Bugs, missing syntax highlighting in some modes, broken extensions—it’s just not stable enough for a deadline.
Outlook
This is only the preview. If Google actually commits to this—fixes the bugs, brings in full extension support, and keeps the Agent Manager workflow—this could be the one.
For now:
- Download it.
- Use it for side projects.
- Play with the aquarium‑game generator.
It’s a glimpse into a future where we do less typing and more managing. That future looks pretty bright—assuming you can get the arrow keys to work.
Are you trusting Google with your code, or are you sticking with VS Code, Cursor, or whatever?