Google Launches Antigravity 2.0: Agent-First Dev Platform at I/O

Google unveiled Antigravity 2.0 at I/O 2026, transforming what was a single IDE feature into a full agentic development suite that lets engineers orchestrate teams of autonomous coding agents from desktop, terminal, or API. The platform shipped Monday and is now generally available to developers worldwide.

Antigravity 2.0 combines four products under one umbrella: a standalone desktop app, a CLI tool, an SDK for embedding agent workflows into your own infrastructure, and Managed Agents in the Gemini API — a hosted runtime where an agent can reason, call tools, and execute code in an isolated Linux sandbox with a single HTTP call.

What Just Shipped

The original Antigravity launched in late 2025 as an in-IDE assistant. Version 2.0 unbundles that experience into a platform. Developers can now switch between three surfaces:

The desktop app is a standalone workspace, not a VS Code extension. It centers on agent orchestration — kick off one agent to refactor a service, another to write the tests, a third to update the docs, and watch them work in parallel inside a single window.

The CLI brings the same harness to the terminal. It's optimized for low overhead and CI workflows, so the same agent you use locally can run in GitHub Actions or a Buildkite pipeline.

The SDK lets developers define custom agent behaviors and deploy them on their own infrastructure. Bring your own model in some configurations, but it's tuned for Gemini 3.5 Pro and Gemini 3.5 Flash.

Managed Agents in the Gemini API is the simplest entry point. One API call spins up an isolated Linux VM where the agent can install packages, run scripts, and persist state across turns. Google handles the sandbox, the cleanup, and the billing.

Why This Matters

Most agent platforms in 2026 are still framework-shaped: LangGraph, Crew, Mastra, and the like. They give you primitives, but you provision your own compute, manage your own state, and write your own glue code to make agents talk to each other.

Antigravity 2.0 sells the opposite pitch: pay Google, get an agent runtime that just works. The bet is that most enterprises don't want to operate agent infrastructure — they want to build features.

This puts Google in direct competition with Anthropic's Claude Agent SDK, which launched in June 2025 and quickly became the standard for production agent workloads, and with OpenAI's Operator. Google's pitch is integration depth: Antigravity agents tie into Google Cloud, Workspace, Drive, Gmail, and BigQuery out of the box.

Key Technical Details

Managed Agents run on isolated Linux environments with internet access, configurable tool allowlists, and a default 30-minute time limit per task that can be extended for long-running jobs. Each agent has access to a virtual filesystem that persists across invocations, so multi-step workflows don't need an external database for state.

The desktop app supports multi-agent orchestration with visual hand-offs. You can wire up a "planner → coder → reviewer" pipeline and the planner agent will route work to specialist agents based on the task.

Pricing is split. The desktop app is free for individual developers. The CLI and SDK use Gemini API token billing — input and output tokens at standard Gemini 3.5 Pro and Flash rates. Managed Agents add a per-second compute charge for the sandbox VM, with a free tier for AI Pro and AI Ultra subscribers.

Google also dropped a limited-time offer: $100 in bonus Antigravity credits for any developer who signs up before May 25.

Industry Impact

The agent-platform race is now four-way: Anthropic Claude Code, OpenAI Codex/Operator, GitHub Copilot Workspace, and Google Antigravity. The first three already have meaningful enterprise traction. Antigravity 2.0 is Google's attempt to leapfrog by bundling tooling that competitors sell separately.

Enterprises building internal developer platforms have been waiting for something exactly like Managed Agents. Self-hosting agent sandboxes is operationally painful — you need container isolation, network egress controls, secret management, and observability. A first-party hosted runtime collapses all of that into a single API endpoint.

For startup founders, the calculus shifts too. Building an "AI engineering assistant" product used to require running your own sandbox infrastructure on Modal, Daytona, or e2b. Managed Agents puts that capability behind a Google API call, which lowers the floor for new agent-native apps.

Expert Perspectives

MarkTechPost called Antigravity 2.0 "the most consequential developer release at I/O 2026" — arguing that the Managed Agents API is the piece that matters most because it commoditizes the hardest part of agent infrastructure.

TechCrunch noted that the CLI version puts Antigravity in direct line-of-sight with Claude Code, which has dominated the terminal-based coding agent space since launch. Whether Google can match Anthropic's iteration speed on the CLI is the open question.

What's Next

Google said Managed Agents will get a vision-input mode in June, allowing agents to take screenshots and reason about UI work. The SDK is also slated to add bring-your-own-model support for select third-party models in the same window.

Developers can sign up at antigravity.google.com, pull the CLI via npm install -g @google/antigravity-cli, or hit the Managed Agents endpoint directly through the Gemini API. The desktop app supports macOS, Linux, and Windows.

Bottom Line

Antigravity 2.0 turns Google from an agent-platform follower into a serious contender. The Managed Agents API in particular changes the economics of building agent-native products — no more standing up your own sandbox infrastructure. If you're shipping anything with autonomous code execution this year, Antigravity is now on the short list of platforms to evaluate.