Onah Sunday.

Posted on May 22

Google Antigravity 2.0 Is the I/O 2026 Announcement You Should Actually Care About

#webdev #ai #devchallenge #googleiochallenge

Google I/O Writing Challenge Submission

This is a submission for the Google I/O Writing Challenge

Everyone is talking about Gemini 3.5 Flash being four times faster than frontier models. Fine. That's real and important.

But if you're a developer who actually ships things — not just watches keynotes — the most structurally significant thing Google announced at I/O 2026 wasn't a model. It wasn't a pricing tier. It wasn't even the intelligent eyewear.

It was Antigravity 2.0.

And most people are going to miss why it matters.

First, a quick rewind

Google launched the original Antigravity in November 2025 alongside Gemini 3. At the time, it was positioned as an AI-native IDE — a Cursor competitor. A heavily modified fork of VS Code, with an agent sidebar, tab completions, and inline commands. An editor view for hands-on coding, plus a Manager surface for spawning and observing multiple agents working asynchronously.

It was interesting. It wasn't a paradigm shift.

Version 2.0 is different. At I/O 2026, Google moved its developer tooling away from IDE-centric assistance and toward multi-agent workflow management as the primary abstraction.

The IDE is now the least interesting part of Antigravity. What they actually shipped is a platform.

What they actually shipped

Here's everything that landed at I/O, in one place:

Antigravity 2.0 desktop app — A new standalone desktop application that enables a full "agent-optimized" user experience. It introduces dynamic subagents for parallelized workflows, scheduled tasks for background automation, and new integrations with Google AI Studio, Android, and Firebase.

Antigravity CLI — Built in Go, snappier and more responsive than Gemini CLI. It fully replaces Gemini CLI, preserving the most critical features: Agent Skills, Hooks, Subagents, and Extensions — now rebranded as Antigravity plugins. Crucially, it shares the same agent harness as the desktop app, meaning all future improvements apply across both surfaces automatically.

Antigravity SDK — Provides programmatic access to the same agent harness that powers Google's own products. Developers can define custom agent behaviors and host them on their own infrastructure.

Managed Agents in the Gemini API — With a single API call, developers can spin up an agent that reasons, uses tools, and executes code in an isolated Linux environment. Three capabilities define this: the agent harness itself, persistent isolated environments where each interaction creates a resumable environment with files and state intact, and custom agent definitions using Markdown skill files.

Enterprise layer — Google Cloud customers can now use Antigravity 2.0 and Antigravity CLI with their Gemini Enterprise Agent Platform project — all agent inference runs via Agent Platform models within a secure cloud boundary, inheriting Google Cloud's data privacy protections.

That's five distinct product surfaces, all sharing the same underlying agent harness. That's not a product update. That's a platform launch.

The part that actually changes how you build

The feature I keep coming back to is Managed Agents. Here's why.

Right now, if you want to add an AI agent to your product — something that reasons, uses tools, and executes code — you have to stand up all the infrastructure yourself. Memory management, tool invocation, isolated execution environments, state persistence across turns. It's a lot of plumbing before you write a line of actual product code.

With Managed Agents, you can spin up an agent that reasons, uses tools, and executes code in an isolated Linux environment with a single API call.

And critically: each interaction creates an environment that can be resumed in follow-up calls with all files and state intact, enabling seamless multi-turn sessions without reinitializing context.

That's persistent, stateful, isolated agent execution — as a managed service. The infrastructure problem is Google's. You just call the API.

If that sounds familiar, it's because it's the same bet Anthropic is making with Claude's computer use and tool use APIs, and the same direction Hermes Agent (Nous Research's open-source agent) took with its persistent gateway model. The difference is Google can embed this directly into the development tools you already use every day.

The Gemini CLI deprecation is the real signal

Gemini CLI will no longer accept new installations for GitHub organizations after June 18, 2026.

This is worth pausing on. Gemini CLI had a massive developer user base. Sunsetting it — even with a migration path — is not a casual decision. It signals that Google is serious about Antigravity as a platform, not a side experiment.

Google is now unifying its developer-facing coding strategy into a single harness across multiple surfaces — editor, terminal, SDK, managed cloud execution. One agent runtime. Multiple entry points. Skills you build once are portable across all of them.

That portability is the core of the bet. Write a Markdown skill file once. Use it in the desktop app, the CLI, the SDK, and Managed Agents in the API. That's how you build a moat around a developer platform — not by having the best model, but by making your ecosystem the easiest place to build agents that work everywhere.

Where it falls short (for now)

I want to be honest about the gaps, because the keynote glossed over them.

Enterprise governance is still coming. Full integration with A2A and Agent Platform governance and security are coming soon — which means enterprise teams at large companies probably can't use this in production yet. Competitors like JetBrains AI Enterprise already ship SOC 2 Type II certification and on-premises deployment options. That gap affects whether security, compliance, and IT teams can approve Antigravity at scale.

The CLI migration isn't frictionless. There won't be 1:1 feature parity right out of the gate. If your team has built workflows and scripts on top of Gemini CLI, you'll need to audit what breaks.

Latency at scale is unproven. The demos showed fast individual agent runs. Multi-agent parallelization at production scale, with real data and real error rates, is a different story. We don't have benchmarks for that yet.

What to try this week

If you want to actually kick the tires, here's where to start:

1. Download Antigravity 2.0 desktop app

Antigravity 2.0 is a GUI app — download it from the official page:

https://antigravity.google/download

Available free for macOS, Windows, and Linux. Sign in with your Google account.

2. Install the Antigravity CLI

The CLI is separate from the desktop app and works in your terminal. These are the real install commands:

# macOS / Linux / WSL
curl -fsSL https://antigravity.google/cli/install.sh | bash

# Windows PowerShell
irm https://antigravity.google/cli/install.ps1 | iex

Note if you're on Windows: I tried the WSL route first and hit a 404 on the wrong URL. The correct path is /cli/install.sh — not /install.sh. Make sure you use the full URL above.

Migrate from Gemini CLI once installed:

antigravity migrate --from gemini-cli

3. Write a custom skill file

Skills are just Markdown — same concept as Hermes Agent's SKILL.md and Claude Code's custom instructions. The syntax is intentionally familiar:

# Skill: pr-reviewer

## Purpose
Review pull requests for security issues, performance regressions,
and style violations.

## Rules
- Flag any hardcoded credentials immediately
- Check for N+1 query patterns in database calls
- Verify error handling on all async operations

Drop it in ~/.gemini/antigravity/skills/ and it's available across all Antigravity surfaces — desktop, CLI, and Managed Agents in the API.

4. Spin up a Managed Agent via the Gemini API

const response = await fetch(
  "https://generativelanguage.googleapis.com/v1beta/agents",
  {
    method: "POST",
    headers: {
      "Content-Type": "application/json",
      "Authorization": `Bearer ${process.env.GEMINI_API_KEY}`
    },
    body: JSON.stringify({
      model: "gemini-3.5-flash",
      system_instruction: "You are a code review agent.",
      tools: ["code_execution", "file_system"],
      persistent_environment: true
    })
  }
);

const agent = await response.json();
// agent.session_id can be resumed across API calls with full file state intact

The bigger picture

Here's the framing I keep returning to: the AI coding tools that win in 2026 won't be the ones with the best autocomplete. They'll be the ones that make agent orchestration the natural next step after writing a function.

Cursor made AI assistance feel native to the editor. Antigravity 2.0 is trying to make agent orchestration feel native to the entire development workflow — editor, terminal, CI/CD, cloud.

Whether Google pulls it off depends on execution. The platform is coherent. The skill portability story is real. The Managed Agents API removes genuine infrastructure friction. But the governance gaps mean enterprises are watching, not deploying. And the Gemini CLI migration will cause friction before it resolves.

The announcement I'd watch most closely isn't the model benchmark or the pricing tier. It's whether Antigravity skill files become as ubiquitous as .gitignore or Dockerfile — shared, versioned, and composable. If that happens, Google wins the agent tooling layer. If it doesn't, Antigravity is a very good IDE.

The keynote made the former seem possible. Only the next six months will tell us which it actually is.