Antigravity: A Practitioner's Field Guide
Antigravity is Google's Gemini running as an autonomous agent inside its own VS Code-like app. It is at its strongest on building visible artifacts — sites, styling, images, deploys — where you can see whether the result is right. The fastest path to value is a small project end-to-end: install, build, ship, in a day.
Building this site, we went from an empty folder to a live editorial build in a day.
The agent set up the framework, generated the templates and styling, configured the deploy pipeline, and pushed the first version to production. We provided the structure and the prose. Antigravity handled everything in between.
That is the cleanest way to understand Antigravity. It is the agent that closes the gap between "I have a rough idea of what I want" and "the thing is online." Use writing-first tools when the output is judgment, structure, or prose. Use Antigravity when the output needs to become a working site, image, or deployed artifact.
Verified February 2026. Both the tool and its pricing change quickly — confirm details against Google's official documentation.
What Antigravity is
Antigravity is Google's AI coding agent. It looks like an editor — the layout is similar to VS Code — but it is a separate application, not a VS Code extension. It runs Google's Gemini models and acts on your project autonomously: editing files, running commands, navigating browsers, generating images.
The important distinction is simple: most chat tools talk. Antigravity acts. It does not just describe what it would do; it does it. Every safety habit in this guide exists because of that line.
Where Antigravity earns its place
| If your work involves | Antigravity is a strong fit |
|---|---|
| Building and shipping websites | Strong with mainstream stacks like Eleventy, Next, Astro, plus deploy targets like Netlify and Vercel |
| Visual or styling work | Live preview inside the editor makes iteration fast |
| Image generation | Built-in tool, no need for a separate service |
| Multi-step technical workflows | The workflow folder lets you codify procedures the agent can replay |
| End-to-end small projects | The task-decomposition UI tracks progress on compound requests |
If your work is mostly long-form writing, planning, or recurring scheduled tasks, OpenClaw is the better default.
Where Antigravity does not fit
| Limitation | Why it matters |
|---|---|
| Sites that require login | Antigravity has no credentials, so authenticated workflows are out of reach |
| Long uninterrupted sessions | Context gets summarized after a few hours; earlier instructions can be forgotten |
| Aesthetic decisions on vibes alone | "Make it look nicer" is one of the most reliable ways to lose a stylesheet |
| Handling secrets | API keys can end up hard-coded into files unless you build the rule against it |
The design issue is worth slowing down on. Antigravity has opinions about how things should look, and a vague request invites them in. Specific instructions — the exact section, color, spacing, font size, and behavior you want — keep the agent inside the change you intended.
Pricing and how to think about it
| Model | Cost shape | Suitable for |
|---|---|---|
| Google AI Studio (free tier) | Free, with rate limits | Trying it out, small projects |
| Gemini Code Assist subscription | Fixed monthly cost (varies by plan) | Day-to-day production use |
| API metered | Pay per use | Operators who understand their consumption pattern |
For this site, we run on the subscription plan. The cost is predictable. The rate limits also act as a guardrail against runaway billing, which is the main risk with metered API setups. The tradeoff is a hard ceiling on throughput at peak load — usually fine, occasionally annoying.
The exact pricing details and tier features change often. Confirm in Google's official documentation before committing.
Three things you will notice in real use
Task decomposition is genuinely good
Compound requests like "draft the article and deploy it" stay manageable. Antigravity breaks them into sub-tasks, runs them in order, and shows progress as it goes. You can see where it is, what it has done, and what is next. You still feel in control while the agent works through the task.
File editing is precise
Edits are presented as diffs. You can see exactly what is changing before applying. For tools that operate on your files, this is the most important thing they can get right, and Antigravity does.
Workflow files turn one-off prompts into reusable operations
Save a repeatable procedure in Antigravity's workflow folder, and the agent can run it again later. For example: publish a new article, deploy with cache busting, or generate a social image. If you repeat a task more than twice, it is worth turning into a workflow.
Three things to watch
Long sessions degrade
After two or three hours, the agent starts losing track of earlier context. The fix is structural: cap sessions at two to three hours, take a break, start fresh. Long sessions are also where the runaway-loop failure mode shows up — we cover that in Three Ways AI Agents Break Your Work.
Command auto-execution can surprise you
Depending on settings and task type, Antigravity may run commands without an explicit confirmation step. The behavior is hard to predict because it varies with the task and the rules in place. The safe default is to list anything you want approval for in AGENTS.md. On a small test build for this site, the agent once ran npm install to add a dependency we had not asked for. No damage was done, but the right reaction was to add an explicit "confirm before installing dependencies" rule and never see it happen again.
Browser tooling is platform-dependent
The browser sub-agent works reliably in Linux environments. On Windows, it may not run at all, depending on the build. If browser automation is part of your workflow, verify on your platform before depending on it.
The fastest path to value: build a small project end-to-end
The strongest first session is not "edit this one file." It is "build something small and publish it."
The pattern that worked for us:
- Tell Antigravity what to build. "Set up an Eleventy site with a homepage, an about page, and a single article. Use a clean editorial style. Get it ready for Netlify deploy."
- Let it run, but watch the diffs. The task-decomposition UI will show what it is doing. Approve scoped changes; redirect when scope drifts.
- Publish the site with one command. Antigravity can walk through the Netlify deployment steps for you, including the parts that normally happen in the command line.
Going from an empty folder to a live editorial site in a day is the experience that makes the tool click. Most people who get it after one weekend project are the people who pick a small enough project to actually finish.
Setup
Prerequisites
- Windows 10 or 11, or macOS
- Node.js v18 or above
- A Google account
Installation
| Step | Action |
|---|---|
| 1 | Download the installer from the official Antigravity site |
| 2 | Launch the application — the layout will look familiar if you have used VS Code |
| 3 | Sign in with your Google account on first launch |
| 4 | Open a project folder and start working |
A note that catches people: Antigravity and Gemini Code Assist are separate products, and Antigravity is not a VS Code extension. It is a standalone application. The visual similarity to VS Code is intentional, but they are different programs.
Three safety essentials
The three rules that prevent most of the incidents we have seen:
| Rule | Why it matters | Time |
|---|---|---|
| Commit to version control before every agent run | Recovery from data loss is one command instead of hours | 10 seconds |
| Watch the command-approval prompts | Catches unintended commands before they execute | Continuous |
| Cap each session at two to three hours | Prevents memory drift, runaway loops, and forgotten earlier instructions | Habit |
The full safety floor is in Three Ways AI Agents Break Your Work and the pre-flight checklist.
How Antigravity compares
| Dimension | Antigravity | OpenClaw | Generic VS Code AI extensions |
|---|---|---|---|
| Underlying model | Gemini (Google) | Claude (Anthropic) | Varies (GPT, Claude, Gemini) |
| File operations | Yes, autonomous | Yes, autonomous | Yes, autonomous |
| Browser control | Yes (Linux primarily) | Yes | Usually no |
| Scheduled tasks | No | Yes | No |
| Standing instructions | AGENTS.md |
SOUL.md |
Tool-specific |
| Task decomposition UI | Yes | No | Step display in some tools |
For the full comparison and decision flow, see OpenClaw vs Antigravity: When to Use Each.
The bottom line
Antigravity is the agent for the part of the work where you can see whether it is right. Sites, styling, images, deploys — anything with a visible result that you can sign off on.
Used well, it can turn a small project from a multi-day slog into a one-day build. Used carelessly, it produces the kind of file-overwrite incidents we keep documenting. The difference is not the tool. It is the rails: version control before every run, command approval as a habit, sessions kept short enough that the agent does not lose its place.
Pick a small project and ship it end to end. That is the fastest way to learn where Antigravity helps and where it needs rails.
Related reading
- OpenClaw vs Antigravity: When to Use Each — head-to-head comparison and decision flow
- OpenClaw: A Practitioner's Field Guide — the companion deep-dive on the Claude-powered alternative
- Three Ways AI Agents Break Your Work — cost, data-loss, and runaway-loop failure modes that apply here
- The Day Antigravity Replaced Our Stylesheet — a concrete incident from a single underspecified prompt
- How to Instruct an AI Agent So It Does What You Asked — the instruction pattern that prevents most of those incidents