Google Antigravity is a free, agent-first development platform from Google that runs autonomous AI agents with a built-in browser, terminal, and code editor — and when you pair it with a simple agent dashboard, your laptop can run a small team of AI workers while you do something else. The platform launched November 18, 2025 alongside the Gemini 3 model family, and as of July 2026 it remains in free public preview with generous rate-limited access to Gemini 3.1 Pro and Gemini 3 Flash models [Wikipedia]. The real power move is not just running agents — it is building a dashboard on top so they run on schedule, report results in one place, and remove the "what should I ask AI today?" problem entirely.
Last verified: 2026-07-30 · Free individual preview · Rate limits reset every 5 hours · Built-in browser automation and multi-agent orchestration
This guide shows you what Antigravity actually does, how its agent architecture works, how to build a lightweight "Mission Control" dashboard for your agents, and the concrete business tasks you can delegate today — none of which require writing code.
What Is Google Antigravity and How Does It Work?
Google Antigravity is an agentic development platform — not just a chatbot that writes code. It is a heavily modified fork of Visual Studio Code that delegates multi-step software development tasks to autonomous AI agents powered by Google's Gemini 3.1 Pro and Gemini 3 Flash models, as well as third-party models like Anthropic Claude Sonnet 4.5 and OpenAI GPT-OSS [Wikipedia]. Google announced it on November 18, 2025, the same day the Gemini 3 model family launched [Google Developers Blog].
The key difference from prior AI coding tools is that Antigravity agents operate across three command surfaces simultaneously: a code editor, an integrated terminal, and a built-in Chromium browser [Cloud Security Alliance]. Unlike an AI plugin that suggests code for you to approve, an Antigravity agent plans, writes files, runs commands, opens websites, clicks buttons, takes screenshots, and reports back what it did.
What Are the Key Features of Antigravity for Agent Work?
The Antigravity feature set is built specifically for orchestrating autonomous agents, not just autocompleting code. Here is what matters:
| Feature | What It Does | Why It Matters for Business |
|---|---|---|
| Manager View | A board showing all your agents at once, each in its own workspace, running in parallel | You see what every agent is doing without opening multiple chat windows |
| Browser-in-the-Loop | Agents control a Chromium browser — navigating, clicking, filling forms, taking screenshots | Agents can test web pages, check competitor sites, and verify UI changes |
| Artifacts | Agents produce reviewable deliverables: task lists, implementation plans, screen recordings, walkthroughs | You watch what the agent did before trusting it — trust through transparency |
| Knowledge Base | Correct an agent once and it remembers; next time it avoids the same mistake | Agents compound learning over sessions instead of starting from zero |
| Plan Comments | Markup the agent's plan like a Google Doc — highlight, comment, and the agent adjusts | You catch a bad plan before execution, not after |
Source: [Antigravity Docs — Best Practices], [Google Developers Blog]
How Much Does Google Antigravity Cost?
As of July 2026, Google Antigravity is free for individual use during its public preview, with an Individual plan listed at $0 per month on Google's pages [Layer3 Labs]. Google has not yet published a stable, standalone paid tier — the current pricing is provisional and the free tier may not reflect the long-term commercial model.
Free does not mean unlimited. Google meters the preview with usage quotas:
- Rate limits reset every 5 hours, so you cannot run agents flat-out all day, but for real batch work the allowance is substantial [AgentDeals].
- Gemini 3 Pro access is included in the free preview; the free tier on the general Gemini API gives 5 RPM and 100 requests per day for Pro-class models, while Flash models get 10 RPM and 1,500 requests per day [AI Prompts Hub].
- Some users reported rate limits kicking in faster than expected at launch; Google has been shipping fixes since [Google AI Developers Forum].
If you need the Gemini API free tier for a broader project, see our guide to Gemini 3.6 Flash for AI Agents: Where It Actually Wins for the model-selection tradeoffs.
How Do You Build an Agent Dashboard ("Agent OS") on Top of Antigravity?
The core insight: Antigravity is great at running agents, but it is still a tool you open, start a task in, and watch. An agent dashboard flips that — it runs above everything, schedules jobs, and puts output in one place. You do not need to buy anything; you can build this with a lightweight web dashboard or even a desktop app that talks to Antigravity's CLI.
The 4-Layer Agent Dashboard Architecture
Layer 4: Dashboard UI → one screen, every agent as a card, status + last output
Layer 3: Scheduler → "run this at 6am, this every 4h, this on-demand"
Layer 2: Agent Runner → calls Antigravity CLI (agy) with prompts, returns artifacts
Layer 1: Antigravity → the engine: browser, terminal, editor, Gemini 3 models
Step 1: Install Antigravity CLI
Antigravity ships as an IDE and a CLI tool. For a dashboard, you want the CLI:
# Install per Google's official guide
# https://antigravity.google/docs/cli/install
The CLI (agy) can run non-interactive one-shot prompts with the -p flag, integrate into shell scripts, and spawn parallel subagents — all programmable from a dashboard backend [Antigravity Docs — Best Practices].
Step 2: Set Up AGENTS.md for Workspace Rules
Create an AGENTS.md file at your workspace root to tell agents your standards, constraints, and prohibited actions. The agent automatically parses these rules on startup and consults them before suggesting changes [Antigravity Docs].
# AGENTS.md
## Rules
- Always verify changes with `npm test` before reporting done
- Never modify files outside src/
- When checking web pages, test on a 375px mobile viewport first
- Report findings as a numbered list with screenshots
Step 3: Build the Scheduler Layer
Your dashboard needs a job queue. Each job is a prompt + a schedule. Use a simple cron-style runner:
| Job | Prompt | Schedule | Output |
|---|---|---|---|
| News Agent | "Summarize the top 10 AI stories from the last 24h" | 6:00 AM daily | Markdown summary |
| Site Checker | "Check all landing pages, flag broken buttons on mobile" | Every 4 hours | Screenshot + issue list |
| Content Drafter | "Draft 3 script outlines from today's top story" | 8:00 AM daily | Markdown outlines |
| Competitor Watch | "Visit these 5 competitor sites, summarize homepage promises" | Mondays 9 AM | Comparison table |
Step 4: Build the Dashboard UI
Each agent becomes a card: green means running, grey means idle, yellow means waiting for your review. Click a card for full history. This removes the starting problem — you do not sit down wondering what to ask AI. The jobs are already set up, they run, and you review.
For a production-grade fleet setup, see our walkthrough on How to Build an AI Agent Fleet on Hermes Agent, which covers the same orchestration pattern for a different agent runtime.
What Are the Best Practices for Prompting Antigravity Agents?
Google's own Antigravity best practices documentation lays out a clear methodology, and it aligns with what security researchers and experienced users recommend [Antigravity Docs], [Antigravity Lab].
Rule 1: Be Specific About the Outcome, Not the Steps
| Bad Prompt | Good Prompt |
|---|---|
| "Improve my landing page" | "On the landing page, a visitor who has never heard of AI automation should understand in 10 seconds what they get and why it matters. Rewrite the top section for that person. Show before and after screenshots." |
| "Check my site" | "Open the homepage on a 375px mobile viewport. Walk through every button. Take a screenshot of each. Flag any button that is below the fold or takes more than 1 tap to reach." |
The agent does not know what "improve" means — it will change fonts. Specify the outcome and the evidence you want.
Rule 2: Explore, Plan, Then Execute
Google's documented workflow is a three-phase approach:
- Exploration: Ask the agent to explain the codebase or page before writing anything.
- Planning: Request an implementation plan. The agent produces an artifact listing targeted files and logic.
- Execution: Once you approve the plan, direct the agent to apply the edits.
This one habit saves hours. As Google's own docs put it: agents do not fail at doing things, they fail at doing the wrong thing very fast [Antigravity Docs].
Rule 3: Establish Verification Loops
Before asking the agent to implement a change, ensure a test suite or verification command exists. Then instruct the agent to run it after its work. The agent watches the output and iterates automatically [Antigravity Docs].
What Can Non-Coders Use Antigravity Agents For?
This is the part most people miss. Antigravity is not just for programmers — it is for anyone whose work involves web pages, documents, research, or repetitive browser tasks. Here are real business tasks you can delegate today:
Customer support mining: "Go through the last 30 customer emails and pull out the five questions people keep asking, then write plain-English answers for each one." That is a support page written for you.
Landing page QA: "Look at the homepage. Find every spot where a new visitor would get confused about what they are joining. Give me five specific fixes." The agent opens the page, walks through it, and sends screenshots with notes.
Competitor research: "Visit these five competitor sites and tell me what they promise on their homepage." The agent opens each one, reads, screenshots, and summarizes — 20 minutes of manual clicking gone.
Social content drafting: "Write 10 short posts explaining what AI automation actually does for a small business in plain words, no jargon." That runs while you are on a call.
Content repurposing: Feed a blog post URL and ask the agent to extract key takeaways, draft a newsletter blurb, and generate social captions — all from the browser without copy-pasting.
For a broader productivity system, see How to Set Up AI Agents for Productivity in 2026: The 6-Step Personal Workflow System.
What Are the Security Risks of Browser-Using AI Agents?
Browser-using agents are a qualitatively different security boundary than a conventional code editor. When an Antigravity agent operates on your machine, it can read and write files, execute shell commands, and interact with browser windows — so the security model matters.
The Pillar Security Prompt Injection Disclosure (Patched Feb 2026)
In April 2026, researchers at Pillar Security disclosed a prompt injection vulnerability in Antigravity that allowed arbitrary code execution by injecting the -X (exec-batch) flag into a file-search tool's pattern parameter — turning a search operation into a binary execution call. This bypassed Antigravity's most restrictive security mode (Secure/Strict Mode) because the search tool was classified as a native system tool and fired before command-level security evaluation [Pillar Security], [Cloud Security Alliance].
The vulnerability was responsibly disclosed and patched by February 28, 2026, and recognized through Google's Vulnerability Reward Program. But the attack class — prompt injection through web content the agent visits — is not specific to Antigravity. Microsoft researchers documented the same pattern across multiple agent frameworks [Microsoft Security Blog].
Permission Gates and the Practical Rule
Google's response is permission gates: the agent only touches browser tabs you explicitly approve. It does not grab your whole browser session [Google Developers Blog].
The practical rule for any agent with browser access: give agents work where a mistake is cheap. Research drafts, checking pages, writing first versions. Do not hand it your payment settings or admin credentials. That is not fear of the technology — it is how you would onboard a new human employee too.
For a deeper look at agent security postures, see our guide on How to Train AI Cybersecurity Defense Models That Actually Out-Think Attackers.
Antigravity vs Other AI Coding Tools: How Does It Compare?
Antigravity is not the only agentic coding platform. Here is how it stacks up on the dimensions that matter for business automation:
| Tool | Free Tier | Browser-in-Loop | Multi-Agent | Best For |
|---|---|---|---|---|
| Google Antigravity | Yes (rate-limited preview) | Yes (built-in Chromium) | Yes (Manager View) | Ambiguous tasks, frontend QA, browser automation |
| Codex CLI (OpenAI) | Requires ChatGPT Plus ($20/mo) | No (needs Playwright MCP) | Single agent | Terminal-native workflows, CI/CD pipelines |
| Cursor | Free + Pro $20/mo | No | Basic | Inline code editing, fast completion |
| Claude Code | API usage billing | No | Standard | Deep reasoning, single-task execution |
Source: [Codex Knowledge Base], [AgentDeals]
Antigravity's edge is the built-in browser and the multi-agent Manager — no other free tool gives you both. For terminal-first automation or deep single-task execution, Codex CLI is the stronger pick. For a free open-source alternative, see our round-up of 11 Free Open-Source AI Tools on GitHub Worth Watching in 2026.
What This Means for You
If you run a small business, produce content, or manage a team: the barrier to AI agents is no longer the software or the price — it is knowing what to point them at. Google Antigravity gives you the engine for free. What you need is a system: a dashboard that holds the jobs, runs them on schedule, and puts the output in front of you.
Start with three agents. One watches your industry and writes a daily summary. One checks your website for issues. One drafts content from the day's best story. Build the dashboard — even a simple Gridsome or Next.js page with cron triggers and agy -p calls works. Run it for a week. You will find that most of your "what should I do with AI" hours were not creativity — they were the starting friction that a dashboard removes.
The agents that fail are the ones given vague prompts with no verification loop. The agents that succeed are given a specific outcome, a plan to approve, and a test to pass. That is the entire game.
FAQ
Q: Is Google Antigravity free to use in 2026? A: Yes, as of July 2026 Google Antigravity is free for individual use during its public preview, with an Individual plan listed at $0/month. Usage is metered with rate limits that reset every 5 hours, but no paid tier has been published yet. Treat the pricing as provisional [Layer3 Labs].
Q: Do you need to know how to code to use Antigravity agents? A: No. Antigravity agents can browse websites, fill forms, take screenshots, and produce reports — none of which requires you to write code. You can ask an agent to check your landing page on mobile, summarize competitor sites, or draft support answers from customer emails without touching a code editor.
Q: How does Antigravity keep agents from running wild in your browser?
A: Antigravity uses permission gates — the agent can only act on browser tabs you explicitly approve. It does not grab your whole browser session. Google also offers three permission modes: request-review (default, prompts before writes), proceed-in-sandbox (safe commands run autonomously), and strict (prompts for all non-read operations) [Antigravity Docs].
Q: Can Antigravity agents be hacked through websites they visit? A: Prompt injection through web content is a real attack class. In February 2026, Pillar Security disclosed a vulnerability where injected instructions in a file-search tool could enable remote code execution, bypassing the most restrictive security mode. Google patched it, but the broader risk — agents being tricked by hidden instructions on pages they browse — applies to all browser-using agents. Keep agents on low-stakes tasks and never give them access to payment or admin systems [Pillar Security].
Q: What models does Antigravity use? A: Antigravity supports Google's Gemini 3.1 Pro and Gemini 3 Flash models as primary engines, and also supports Anthropic Claude Sonnet 4.5 and OpenAI GPT-OSS as third-party options. Per-mission model selection lets you route cheaper tasks to Flash models while reserving Pro for the final pass [Wikipedia].
Q: How is an "agent dashboard" different from just opening Antigravity? A: Antigravity is a tool you open, start a task in, and watch. An agent dashboard sits above it — it schedules jobs ("run this at 6 AM, this every 4 hours"), runs them automatically, and shows you all outputs in one screen. It removes the "starting problem" where you have to decide what to ask AI. The jobs are pre-set, they run, you review results.
Discussion
0 comments