Google Antigravity · AI Agents · 2026

Google Antigravity: the 4 products (2.0, CLI, IDE & SDK)

Not a CLI. Not an IDE. Four products on the same engine — Google's shared "agent harness". I tested it for a full week.

What actually happened

Google didn't launch one product. It launched four — and all four run on the same engine.

At Google I/O, Google said it already processes 500 billion tokens a day internally with its developers on this platform. Not a PoC: real production. The headlines talked about the "new AI IDE" or the "Gemini CLI replacement". But the full story is the suite.

            ANTIGRAVITY HARNESS  (shared engine)
                        │
   ┌────────────┬───────┴───────┬────────────┐
   ▼            ▼               ▼            ▼
Antigravity  Antigravity    Antigravity  Antigravity
2.0          CLI            IDE          SDK
(desktop)    (terminal)     (editor)     (Python)

The 4 products

Almost every video covers only the IDE or only the CLI. There are four, not one — each optimized for a different workflow.

1

Antigravity 2.0

Desktop

A standalone desktop app — and Google is explicit: "there is no IDE". It's a command center: your job is to direct agents, not write code. It ships Projects, Scheduled Tasks, voice dictation and native Git worktrees.

2

Antigravity CLI

Terminal

The terminal version of the same engine. The binary is called agy (not "antigravity" — that trips people up). Keyboard-first, ideal for SSH and tmux. Shares configuration, permissions and MCP servers with 2.0.

3

Antigravity IDE

Editor

What Google launched first, in November 2025. A full code editor (Open VSX fork, VS Code-like) with autocomplete and an agent in the side panel. 2.0 did NOT replace it: Google recommends "dual-wielding" them.

4

Antigravity SDK

Python

A Python library: pip install google-antigravity. You embed Google's own agent into your app, script or service. And it's the only product in the suite that lets you build real Layer 2.

The shared "agent harness"

Google says every surface runs on a "shared agent harness". That term isn't a Google invention: the whole industry has been adopting it since November 2025.

1

Nov 26 2025 · Anthropic

"Effective Harnesses for Long-Running Agents" — the original harness pattern.

2

Feb 5 2026 · Mitchell Hashimoto

Terraform's creator: "engineer the harness". Coins the term.

3

Feb 11 2026 · OpenAI

"Harness Engineering: Leveraging Codex" — 1M lines, zero written by humans.

4

Apr 2 2026 · Martin Fowler

Formal taxonomy: Guides + Sensors. When Fowler orders it, the industry adopts it.

5

Apr 29 2026 · Karpathy

Vibe coding dead → agentic engineering.

6

May 19 2026 · Google

"shared agent harness" — Antigravity's 4 products run on the same engine.

One important nuance: the "Antigravity harness" is Google's specific runtime — one implementation of the discipline, not the whole discipline. Watch Harness Engineering ›

Layer 1 vs Layer 2: what it enforces and what it doesn't

The difference that defines when to use Antigravity as-is, and when you need to build your own agent.

Layer 1 — by convention

The agent can do the dangerous action, but asks you first (approval gate, sandbox, permissions). It trusts you to approve well. Antigravity's is brutal: native sandbox (nsjail, sandbox-exec, AppContainer), regex permissions with Deny > Ask > Allow precedence, and 9 lifecycle hooks.

Layer 2 — by design

The dangerous capability doesn't exist in the code. The agent can't delete a DNS record because the function isn't registered. It doesn't depend on your judgment. In the suite, only the SDK allows real Layer 2: you register only the tools you choose.

Technical verdict, neutral and honest

Antigravity has a brutal harness in the product sense — its Layer 1 is one of the best on the market. But if you want real architectural restriction (Layer 2), you have to build it yourself with the SDK. Google doesn't enforce it by default. This isn't a criticism: it's a blueprint of the product.

5 models on the Free tier

No extra API key, no credit card. Switch them live with /model.

Gemini 3.5 FlashGemini 3.1 ProClaude Sonnet 4.6Claude Opus 4.6GPT-OSS 120B

Yes — Claude Sonnet 4.6, Claude Opus 4.6 and GPT-OSS 120B, free inside Antigravity. The default is Gemini 3.5 Flash, the one Google optimized (per the I/O 2026 blog, 12× faster on Antigravity):

76.2%

Terminal-Bench 2.1

1656 Elo

GDPval-AA

83.6%

MCP Atlas

84.2%

CharXiv Reasoning

The CLI ↔ 2.0 bridge: one session, two surfaces

"shared agent harness" in action: start on the desktop, continue in the terminal without losing context. This is what no other video covers with its caveats.

2.0 → CLI · clean

Launch the CLI with /resume, press Tab to switch to the 2.0 conversations, pick the session and continue where you were. This is the direction I recommend for production.

CLI → 2.0 · unstable today

It works, but you must do 3 things in order: close the CLI with /exit (the SQLite uses Write-Ahead Log), align the Project's workspace with the CLI's folder, and refresh the @conversation dropdown. There are reported import bugs.

Bonus: /fork clones the entire conversation to try alternative approaches without losing the main branch.

The SDK: your own Layer 2

The piece almost nobody shows, and the one that gives real control. With pip install google-antigravity and about 15 lines you get an agent on the same engine Google uses internally.

Replace the defaults entirely

The SDK lets you replace the toolset with your own. If you don't include run_command, the agent literally doesn't have that tool. Not an approval gate: absence of capability. Real Layer 2.

9 lifecycle hooks

From session start to context compaction. Each hook with three modes: inspect (logging), decide (approve/block) and transform (modify in-flight data). Native observability for enterprise and compliance.

Roadmap: Gemma

Remote harness on Cloud Run, TypeScript and Go, and the biggest one: Gemma integration — the full harness on open models you control and fine-tune. Key for companies that can't send data to Google.

The drama: the Claude Code clause

The part almost nobody covers, and that you need to know before building a mixed stack in production.

Verbatim · antigravity.google/docs/faq

"Using third party software, tools, or services — e.g. Claude Code, OpenClaw, OpenCode — to access Antigravity is a violation of our Terms of Service. Such actions may be grounds for suspension or termination of your account."

If you use Claude Code passing it off as a legitimate Antigravity session (proxying requests with your login), Google can terminate your account. The solution Google itself recommends in the same FAQ: use a Vertex AI or AI Studio API key.

The legal mixed setup (per Google)

Antigravity 2.0  ─► Google login   ─► Antigravity harness
Antigravity CLI  ─► Google login   ─► Antigravity harness
Claude Code      ─► Vertex API key ─► Gemini direct (NOT Antigravity)
Codex local      ─► Vertex API key ─► Gemini direct (NOT Antigravity)

Google builds Antigravity as a closed ecosystem on top (your login + the 4 products) and an open API underneath (Vertex/AI Studio for everything else). Two planes, two contracts.

Migrating from Gemini CLI

Gemini CLI stopped serving requests on June 18, 2026. Migration is mandatory — but trivial, ten minutes.

agy plugin import gemini   # converts skills, commands and MCP servers

Config

Moves to ~/.gemini/antigravity-cli/. The root name .gemini didn't change — no broken retro-compat.

Skills

From .gemini/skills/ to .agents/skills/ — the open AGENTS.md standard Codex already uses.

MCP

The url/httpUrl key becomes serverUrl on remote servers. Local (stdio) ones don't change.

The scale: subagents and Scheduled Tasks

What changes the mental model: from "one agent writing code" to dozens coordinating in parallel.

Google's I/O demo

93 subagents in parallel built the core of an operating system from scratch — in 12 hours.

93

Subagents in parallel

15,000

Model requests

2.6 B

Tokens processed

"This was not possible on Gemini 3.1 Pro." — Google I/O 2026

Shortcuts to skip micro-approving

Alt+J teleports you to the subagent asking for approval; Ctrl+K approves without leaving the main prompt. On big refactors, they save minutes per session.

Scheduled Tasks · 2.0 only

Cron jobs for agents: "every weekday at 9am review my PRs and summarize which are ready to merge". The agent has its own agenda while you sleep. Still new, 2.0 only.

When to use each (and when not)

Honest trade-offs. When a product is NOT for you, I'll tell you too.

Product When YES
Antigravity 2.0 Orchestrate several agents in parallel from a visual interface, Scheduled Tasks, the "command center" model.
Antigravity CLI SSH/tmux work, terminal-first, or migrating from Gemini CLI.
Antigravity IDE You come from Cursor or Windsurf and edit a lot of code manually with strong autocomplete.
Antigravity SDK Build custom agents for production with real Layer 2, host them on your own infra, or embed the harness in your product.

When NOT

Sources

All quotes are verbatim from the official Antigravity documentation.

Google · antigravity.google · 2026

Antigravity — product & docs

The product page and official documentation: the 4 products, the "shared agent harness" and the toolset.

Google · Blog · May 2026

Google I/O 2026

500 billion internal tokens/day, Gemini 3.5 Flash benchmarks, and the 93 subagents that built an OS in 12 hours.

Google · Blog · 2026

Introducing Antigravity 2.0

The desktop app with no IDE ("there is no IDE") and the recommendation to use it alongside your editor.

Google · Blog · 2026

Introducing the Antigravity SDK

pip install google-antigravity, the 9 lifecycle hooks and the roadmap (Gemma, TypeScript, Go, remote harness).

Google · Docs · 2026

FAQ — Terms of Service

The literal clause: using Claude Code, OpenClaw or OpenCode with your Antigravity login violates the Terms of Service.

Related videos

Antigravity connects directly with the channel's series on harness, loops and agents in production.

Harness Engineering

The base concept. The "shared agent harness" is Google's implementation of this discipline. Agent = Model + Harness.

Loop Engineering

How the agent's work repeats over time — exactly what Scheduled Tasks and parallel subagents do.

MCP in Production

The connectivity layer. Antigravity shares MCP servers between 2.0 and the CLI; here you see MCP at scale.

ADK + A2A on Cloud Run

The real Layer 2 example: the delete_dns_record function doesn't exist in the agent's code. It can't call it.

Claude Agent Teams

Subagents splitting the work — the pattern Antigravity takes to 93 agents in parallel.

Google ADK

How to build an agent with scoped tools and governance — the principle behind the SDK's Layer 2.

Frequently asked questions

The essentials about Google Antigravity.

What is Google Antigravity?

+

It's Google's agentic development platform. And the key: it's not one product, it's four — Antigravity 2.0 (desktop), Antigravity CLI (terminal), Antigravity IDE (editor) and Antigravity SDK (Python library) — all running on a shared "agent harness" co-trained with the Gemini models.

How do I install and use Antigravity?

+

Antigravity 2.0 and the IDE are desktop apps: download them from antigravity.google/download (macOS, Windows, Linux). The CLI and SDK go through the terminal: the CLI via a script (curl ... | bash) and its binary is called agy; the SDK via pip install google-antigravity. On first run the CLI asks for theme, render mode and workspace trust.

Can I use Claude Code with my Antigravity login?

+

No. Google's official FAQ states verbatim that using third-party software (it names Claude Code, OpenClaw and OpenCode) to access Antigravity with your login violates the Terms of Service and may be grounds for suspension or termination of your account. Google's own recommended solution: use a Vertex AI or AI Studio API key, which points at Gemini directly, not at your Antigravity login.

Is Antigravity free? Which models does it include?

+

It has a Free tier with access to five models without an extra API key: Gemini 3.5 Flash, Gemini 3.1 Pro, Claude Sonnet 4.6, Claude Opus 4.6 and GPT-OSS 120B (with several reasoning levels). The Free tier's rate limits are tight and the Claude/GPT-OSS models have separate quota from the Gemini ones; for heavy use you need the Pro plan.

What's the difference between Antigravity 2.0, CLI, IDE and SDK?

+

2.0 is a desktop "command center" app for directing agents (not an editor). The CLI is the terminal version (binary agy), keyboard-first. The IDE is a full code editor (Open VSX fork, VS Code-like). The SDK is a Python library to build your own agents. All four share the same engine; the surface changes.

How do I migrate from Gemini CLI to Antigravity CLI?

+

With one command: agy plugin import gemini, which converts your skills, custom commands and MCP servers. There are three changes: config moves to ~/.gemini/antigravity-cli/, skills move to .agents/ (the AGENTS.md standard), and the MCP schema changes the url/httpUrl key to serverUrl. Gemini CLI stopped serving requests on June 18, 2026 — migration is mandatory.

Does Antigravity provide Layer 2 security (architectural restriction)?

+

Not by default. Antigravity ships a very strong Layer 1 — native sandbox (nsjail, sandbox-exec, AppContainer), regex permissions with Deny > Ask > Allow precedence, and approval gates — but the agent does have the dangerous tools. Real Layer 2 (where the capability doesn't exist in the code) you only get by building your agent with the SDK and registering only the tools you need.

Community

Full Antigravity course in Agentic Engineers

The step-by-step migration from Gemini CLI, the SDK with your own Layer 2, and real repos to practice. Free community access; the full course is in the Premium tier.

Join Agentic Engineers →

YouTube channel

@NicolasNeiraGarcia

ADK · A2A · Claude Code · Automation · Infrastructure

Subscribe ›