Google Antigravity Review: Is Google's Agent-First Coding Platform Worth the Switch?

For most of 2025, "AI coding tool" meant Cursor with a different paint job. Then in November Google quietly shipped something that didn't look like Cursor at all — a desktop app where AI agents are first-class citizens and the code editor is almost a secondary surface. Six months later, at Google I/O 2026, they relaunched it as Antigravity 2.0 with a standalone app, a CLI, an SDK, and enterprise hooks. We've been using it daily for the past month. Here's the honest review.

What Antigravity actually is

Antigravity is not an IDE with a chat panel. It's a development platform built around the assumption that you will increasingly delegate end-to-end software tasks to autonomous agents — and that you need a place to plan, watch, and verify what those agents do.

The product has two main surfaces:

• Editor View — a familiar AI-powered code editor (it's based on VS Code under the hood, with tab completions and inline commands). This is where you still work when you want to be hands-on.
• Manager Surface — a dedicated dashboard for spawning, orchestrating, and observing multiple agents that work asynchronously across separate workspaces. Each agent has its own editor, terminal, and embedded browser. You give it a goal, it goes off and works.

The Manager Surface is the part that actually changes how you work. You stop context-switching every 20 seconds.

What's new in Antigravity 2.0

The November 2025 launch shipped the IDE. The May 2026 I/O update reframed Antigravity as a multi-surface suite. The headline additions:

• Standalone desktop application — separate from the original Antigravity IDE, fully optimized around agent orchestration rather than file editing.
• Antigravity CLI — a lightweight terminal-first surface where you can spin up agents without ever opening a graphical interface. Replaces the deprecated Gemini CLI, which Google is sunsetting on June 18, 2026.
• Antigravity SDK — programmatic access to the same agent harness that powers Google's own products. Lets you define custom agents and host them on your own infrastructure.
• Managed Agents in the Gemini API — agents you don't have to host yourself. Google runs them.
• Dynamic subagents — your main agent can now spawn parallel subagents for sub-tasks instead of doing everything sequentially. Big speed win on multi-file refactors.
• Scheduled tasks — agents that run on a cron-like schedule for background automation (overnight test suites, dependency updates, log triage).
• Enterprise tier via Gemini Enterprise Agent Platform — directly connectable to Google Cloud projects.

Models: not just Gemini

This was the surprise. Antigravity is not a Gemini-only product. At launch it offered Gemini 3 Pro, then 2.0 brought generous rate limits on the new Gemini 3.5 Flash (Google's reported benchmark on Terminal Bench 2.1 is 76.2%, beating Gemini 3.1 Pro at 70.3%). But it also fully supports:

• Anthropic's Claude Sonnet 4.5 — most users we know default to this for complex multi-file work.
• OpenAI's GPT-OSS — for users who want an open-weight option.

Being able to switch models mid-task without leaving the app is a real productivity unlock. Gemini 3.5 Flash for fast iteration, Claude Sonnet 4.5 when you need careful reasoning over a large codebase.

How it actually feels day-to-day

We gave Antigravity three real tasks during testing:

1. "Add OAuth login to this existing Express app" — Antigravity planned the steps, modified the relevant files, ran the dev server in its embedded terminal, then opened its embedded browser to test the login flow itself. It caught a missing redirect URI on its second attempt without prompting. Total time: ~6 minutes, mostly waiting.
2. "Migrate this codebase from CRA to Vite" — Took longer (about 25 minutes) because the agent had to handle config conflicts. The Manager Surface log was genuinely useful here — we could see what it was trying, abort once when it went down a bad path, and redirect it.
3. "Find and fix the memory leak in this Node.js worker" — Mixed result. It correctly identified the leaking module but proposed a fix that papered over the symptom. Required a second prompt to get the actual root cause. Still faster than doing it manually.

The honest takeaway: Antigravity is not magic. It still hallucinates, still needs supervision, still occasionally goes down rabbit holes. But the Manager Surface gives you the visibility to catch those moments in a way that chat-panel IDEs simply do not.

Antigravity vs Cursor vs Claude Code

If you live in your terminal and want minimal abstraction, Claude Code still wins. If you want the fastest inline AI autocomplete on the planet, Cursor is still ahead. But for tasks that span multiple files, multiple tools, and multiple minutes, Antigravity's Manager Surface is the most thoughtful UX anyone has shipped so far.

The catch

A few things to know before you switch:

• Token-intensive — agent workflows burn through tokens much faster than chat. Google has reportedly introduced a new $100/mo AI Ultra tier partly because of this. The free preview has generous limits, but heavy users will hit them.
• Still preview — features change, APIs change, occasional crashes. Don't bet a production pipeline on it yet.
• Chrome is a hard dependency — agents use Chrome for the embedded browser. Firefox-only setups won't work.
• Workspace permission model is broad — agents can read and write anything in their assigned folder. Be careful about which folder you point them at, especially with credentials or secrets nearby.

Who should use it

The bottom line

Antigravity isn't perfect, but it is the first AI coding tool that feels designed for a world where you trust agents to do real work. The bet Google is making is that within 18 months, "how many agents you can orchestrate in parallel" will matter more than "how fast your autocomplete is." On current evidence, that bet looks correct.

It's free. It's cross-platform. It supports the models you already use. The only reason not to try it is inertia.