Google Antigravity 2.0 is a standalone desktop application that orchestrates autonomous AI agent teams — and with the right configuration, it functions as a genuine agent operating system for your work. Google's own team proved this by having 93 sub-agents build a working operating system from a single prompt in 12 hours for $916.92. The same building blocks — projects, memory files, skills, scheduled tasks, and hooks — are available to every user today.
TL;DR:
- Antigravity 2.0 is a standalone desktop app (Mac, Windows, Linux) with no built-in code editor — it is built around agent orchestration, not file editing
- Google's OS-from-scratch demo used 93 sub-agents, 15,314 model calls, and 339M+ input tokens on Gemini 3.5 Flash for $916.92
- Five building blocks turn the empty chat app into an agent OS: Projects, Memory (AGENTS.md), Skills, Scheduled Tasks, and Hooks
- A 4-day setup plan gets you from zero to autonomous agent workflows
- The
/teamwork-previewmulti-agent feature requires the $200/month AI Ultra plan; the free tier gives 20 agent requests/day
What is Google Antigravity 2.0?
Antigravity 2.0 is a standalone desktop application for macOS, Linux, and Windows that Google launched on May 19, 2026, at Google I/O. Unlike Antigravity 1.0, which was an AI-powered IDE (code editor), version 2.0 has no IDE at all. You open it and see a chat window with an agent. The code editor was removed deliberately — Google's team reasoned that agents deserve a dedicated workspace, not a sidebar in an IDE. (Source: Google Antigravity Blog)
The app runs on Gemini models (currently Gemini 3.5 Flash) and lets you interact with agents both synchronously (chat-based, human-in-the-loop) and asynchronously (fire-and-forget, autonomous). The key architectural shift is that the agent system — not the code editor — is the product. Multi-agent teams, scheduled tasks, native voice, and project-level configuration are the core features. (Source: TechCrunch)
What did Google's 93-agent operating system demo actually prove?
Google tasked a team of 93 sub-agents with building a working operating system from scratch — kernel, process and memory management, a filesystem, and video and keyboard drivers — from a single prompt. The result ran FreeDoom. The verified numbers from Google's own engineering blog:
| Metric | Value |
|---|---|
| Sub-agents | 93 |
| Model calls | 15,314 |
| Input tokens | 339M+ (2.6B total with cache reads, output, thinking) |
| Wall-clock time | 12 hours |
| Cost at API pricing | $916.92 |
| Model | Gemini 3.5 Flash |
| Previous model attempt | Gemini 3.1 Pro — could not do it |
(Source: Google Antigravity Engineering Blog)
Google noted two things that matter for anyone using Antigravity 2.0: (1) they did not use a secret internal version — the demo ran on the same features available to every user (sub-agents, scheduled tasks, hooks, projects), and (2) the agents initially "cheated" by referencing conversations from past runs, so Google implemented anti-cheating guardrails to ensure genuine zero-to-one autonomy. That anti-cheating mechanism is the smartest trick in the whole system, and the same hooks system is available to you.
Independent researchers at AI Snake Oil (Sayash Kapoor and Arvind Narayanan, Princeton) noted that the "single prompt" was actually many thousands of lines long, and Google has not released the prompt, code, or logs — so independent evaluation is limited. The OS lacks floating-point math, hardware acceleration, multi-threading, sandboxing, and JIT compilation. It is a proof-of-capability, not production software. (Source: AI as Normal Technology)
How do you turn Antigravity 2.0 into an agent operating system?
An agent operating system means your AI agents have persistent memory, defined projects, reusable skills, autonomous schedules, and safety guardrails — instead of starting from zero every time you type a prompt. Here are the five building blocks, in the order you should set them up.
1. Projects: give your agent the right context
In Antigravity 1.0, your agent was locked to a single code repository. Version 2.0 replaced this with Projects — configurations that can point to multiple folders simultaneously. A project gets its own settings, permissions, and agent configuration. (Source: Antigravity Docs — Projects)
This is the change that makes Antigravity usable for non-coders. A project called "content" can point to your notes folder, your brand assets folder, and your website folder all at once. The agent sees all of them together. You are not limited to code repositories — a project can be a folder of Word documents and spreadsheets.
How to set it up:
- Open Antigravity 2.0 and create a new Project
- Add the folders the agent needs access to (notes, brand files, website, etc.)
- Configure permissions for that project — restrict what the agent can do
- Set project-level agent settings (model, behavior)
2. Memory: teach the agent your business with AGENTS.md
Every agent in Antigravity reads rules files from your project at the start of each session. The primary file is AGENTS.md — a plain-text Markdown file you place in your project root that explains who you are, what you sell, who your customers are, how you talk, and what you never say. Antigravity has supported AGENTS.md since v1.20.3 (March 2026). (Source: Agentpedia — AGENTS.md Guide)
The alternative file is GEMINI.md, which is Antigravity-specific and takes precedence over AGENTS.md when rules conflict. For teams using multiple AI tools (Cursor, Claude Code, Antigravity), put shared rules in AGENTS.md and Antigravity-specific overrides in GEMINI.md. (Source: Agentpedia — Rules Guide)
Without a memory file, you are prompting a stranger every morning. With one, the agent starts every session already knowing your business context.
What to put in your AGENTS.md:
- Who you are and what your business does
- Who your customers are and how you talk to them
- What you sell (and what you do not)
- Coding or writing conventions
- Safety guardrails (what the agent must never do autonomously)
Your memory file should live in plain text on your machine, not in a cloud tool — you own it, it stays private, and it works offline. A half-page beats nothing. The file is read at session start and applies throughout the agent's work.
3. Skills: write your process once, reuse forever
Skills are folders containing instructions that agents can call on demand. You write down a process once — and the agent can invoke that skill whenever it is needed. Google shipped a built-in skill called the "Antigravity Guide" in June 2026: if you ask the agent how Antigravity works, it reads its own manual instead of guessing. (Source: Antigravity Docs — Skills)
Think about what this means practically. Every process you have ever explained to a person twice — three times, ten times — becomes a file you write once. A skill might be "write an SEO blog post in our exact format with our exact structure" or "generate a weekly sales report from the spreadsheet in the reports folder."
How to create a skill:
- Create a folder under
.agent/skills/(or the global skills directory) - Write a
SKILL.mdfile with the instructions - Include any supporting files (templates, scripts, examples)
- The agent automatically discovers and can invoke the skill when the task matches
Skills are portable — the same AGENTS.md and skill folder format works across Antigravity, Cursor, and Claude Code, so your investment in writing skills compounds across tools.
4. Scheduled Tasks: make agents work while you sleep
Antigravity 2.0 added scheduled tasks — triggered with the /schedule command. You tell an agent to run on a timer: once, tomorrow at 9:00 AM, every morning, or every Monday. Google's framing: "you no longer need to manually invoke every agent." (Source: Antigravity Docs)
This is the actual difference between using AI and having AI work for you. Most people sit down, open a chat, type something, wait, copy the answer, close the laptop. Nothing runs while they sleep. With scheduled tasks, you set it up once and it happens on schedule.
Practical scheduled task patterns:
- A daily content review agent that checks your drafts folder every morning
- A weekly report generator that pulls metrics on Monday at 8 AM
- A monitoring agent that checks your website uptime every hour
- A content repurposing agent that turns your latest blog post into social snippets
Critical safety note: Scheduled tasks can burn through your API credits while you are not watching. Always set cooldowns, use locked directories for output, and keep a deactivation switch that requires no code modification. Google explicitly warns that the /teamwork-preview multi-agent feature will exhaust your entire weekly quota within a couple of tasks. (Source: Google Antigravity Blog)
5. Hooks: stop agents from faking their work
Hooks are event-triggered rules that fire automatically when certain things happen — before a file is saved, after code is generated, when an agent claims it finished a task. Google used hooks in the OS demo to stop agents from cheating: when the agents referenced past conversation data instead of building from scratch, the hook detected it and killed the run. (Source: Antigravity Docs — Hooks)
Hooks are the safety layer that makes autonomous agent work trustworthy. Without them, an agent can claim success without actually completing the task — and you would not know until you checked.
What hooks should you set up?
- A pre-save hook that runs a format check and secret scan
- A post-completion hook that runs tests before marking a task done
- A cheating-detection hook that checks whether the agent is referencing stale conversation data
- A guardrail hook that blocks certain file paths or directories from being modified
Hooks must be fast and deterministic. Do not use hooks for long-running builds, end-to-end tests, or deployments — those should be scheduled tasks or sub-agents. Hook failure should prevent the next step and hand the exit code to the agent manager. (Source: KnightLi Blog)
How much does Antigravity 2.0 cost?
Antigravity 2.0 has a free tier and paid tiers through Google AI subscriptions:
| Plan | Price/month | Usage limit | Key features |
|---|---|---|---|
| Free | $0 | 20 agent requests/day | Gemini Flash model, tab completions |
| AI Pro | $20 | ~100 agent requests/day | More capable models, weekly quota refresh |
| AI Ultra (new) | $100 | 5x Pro limits | Antigravity 2.0 priority, 20TB storage, Gemini Spark beta |
| AI Ultra (top) | $200 | 20x Pro limits | /teamwork-preview multi-agent, 25,000 AI credits/month |
The multi-agent /teamwork-preview feature — the one Google used for the 93-agent OS demo — is gated to the $200/month AI Ultra plan. Google warns that even on this plan, a single multi-agent task can exhaust your entire weekly quota. (Source: Google AI Blog, Source: Google Antigravity Blog)
Pricing and limits are volatile — re-check before committing to volume work.
What is the 4-day setup plan for an agent operating system?
You do not build the whole system on day one. The people who get stuck are always the ones trying to do everything at once. Here is a day-by-day plan — one hour per day:
Day 1: Build your memory file. Write an AGENTS.md that explains who you are, what you sell, who you serve, and how you talk. A half-page beats nothing. The agent now starts every session knowing your business instead of being a stranger.
Day 2: Create one project and point it at the folders that matter. Set up a single project in Antigravity 2.0 that includes the directories the agent needs — your documents, your website files, your brand assets. Configure permissions so the agent can read what it needs but cannot touch what it should not.
Day 3: Write one skill for the job you do most often. Whatever task you find yourself explaining to people repeatedly — writing blog posts, generating reports, reviewing code — turn it into a skill folder with a SKILL.md file. The agent can now call that skill on demand.
Day 4: Schedule it. Use the /schedule command to set the agent to run that skill on a timer. Now the work happens without you. Start with one scheduled task — not ten.
One automation per week. Add a second skill in week two. A third scheduled task in week three. The system grows incrementally, and each piece is tested before the next one is added.
Antigravity 2.0 vs other agent platforms: how does it compare?
| Feature | Antigravity 2.0 | Claude Code | OpenAI Codex | Cursor |
|---|---|---|---|---|
| Core model | Gemini 3.5 Flash | Claude Opus/Sonnet | GPT-5.5 | Multiple |
| Multi-agent teams | Yes (93-agent demo) | Subagents | Limited | Composer |
| Scheduled tasks | Built-in (/schedule) | Third-party skills | No | No |
| Memory files | AGENTS.md + GEMINI.md | CLAUDE.md | — | .cursorrules |
| Skills system | Yes (portable) | Yes | — | Yes |
| Native voice | Yes (live transcription) | No | No | No |
| IDE included | No (standalone app) | No (CLI) | No (CLI) | Yes (full IDE) |
| Free tier | 20 requests/day | — | — | Limited |
| Top plan | $200/month | $100/month (Max) | $200/month (Pro) | $20+/month |
Antigravity 2.0's genuine differentiators are native voice, built-in scheduled tasks, and deep Google ecosystem integration. The subagent pattern exists in Claude Code too; scheduled tasks can be added to Claude via third-party skills. The decision comes down to which model ecosystem you are already in and whether the Google integration matters for your workflow. (Source: MCP.Directory)
What this means for you
If you are a small business owner, freelancer, or builder: the gap between "typing questions into a chat box one at a time" and "having a system that runs while you sleep" is now five configuration steps. You do not need the $200/month plan or 93 sub-agents. The free tier gives you 20 agent requests per day, and the building blocks (projects, memory, skills, scheduled tasks, hooks) work on every plan.
Start with memory. An AGENTS.md file that explains your business is the single highest-leverage thing you can do — it transforms every future agent interaction from "prompting a stranger" to "instructing someone who already knows your context." Then add one project, one skill, and one scheduled task over four days. That is your agent operating system.
If you want to explore other agent OS approaches, see our guide on how to build an AI agent operating system with open-source tools, or learn about automated lead generation with AI agents that runs on a schedule. For a broader view of AI automation patterns, our SENSE framework for AI adoption provides a 30-day plan for small businesses.
FAQ
Q: Is Antigravity 2.0 free to use?
A: Yes, there is a free tier that gives you 20 agent requests per day using the Gemini Flash model, with no credit card required. Paid plans start at $20/month (AI Pro) and go up to $200/month (AI Ultra) for the multi-agent /teamwork-preview feature.
Q: Do I need to know how to code to use Antigravity 2.0? A: No. Antigravity 2.0 has no code editor — it is a standalone chat-based agent app. Projects can point to folders of Word documents and spreadsheets, not just code. The AGENTS.md memory file and skills are written in plain Markdown, not code.
Q: What is the difference between AGENTS.md and GEMINI.md? A: AGENTS.md is a cross-tool rules file read by Antigravity, Cursor, and Claude Code. GEMINI.md is Antigravity-specific. When both exist, GEMINI.md takes precedence for conflicting rules. Use AGENTS.md for shared rules and GEMINI.md for Antigravity-only overrides.
Q: Can agents in Antigravity 2.0 run on a schedule without me watching?
A: Yes. The /schedule command lets you set agents to run once, at a specific time, or on a recurring schedule. Google warns that scheduled multi-agent tasks can exhaust your weekly quota quickly — monitor usage and set cooldowns.
Q: What model does Antigravity 2.0 use? A: Antigravity 2.0 runs on Gemini models, currently Gemini 3.5 Flash by default. Google's OS demo used Gemini 3.5 Flash and noted that the older Gemini 3.1 Pro could not complete the same task. Gemini 3.5 Pro was announced for release in June 2026.
Q: What is the /teamwork-preview feature?
A: /teamwork-preview is the multi-agent team feature Google used for the 93-agent OS demo. It is available as a research preview to AI Ultra subscribers on the $200/month plan. Google recommends using it with Gemini 3.5 Flash to avoid hefty bills, and warns it will exhaust your weekly quota within a couple of tasks.
Discussion
0 comments