Antigravity, eleven months in

The short version

Google launched Antigravity on 18 November 2025, the same day as Gemini 3. Back then it was one thing: an agent-first IDE built on a VS Code fork, made largely by the Windsurf people Google brought over in its roughly $2.4 billion licensing deal the summer before.

It is not one thing anymore. At Google I/O on 19 May 2026 it became a family. There is the Antigravity 2.0 desktop app, where you run several agents in parallel and schedule work in the background. There is agy, a terminal agent that replaced Gemini CLI. There is a Python SDK you install with pip install google-antigravity. And the IDE side now ships as extensions for VS Code, Visual Studio, JetBrains, Zed and Xcode, alongside the standalone Antigravity IDE.

Below I go through what improved, what pushed it, how it looks and feels next to the other tools I use or watch, and where I think it still falls short.

01What got better

At launch Antigravity already had two views: an editor with an agent panel, and a Manager view for running agents across workspaces. Early reviews liked the speed and complained about stability. The bigger changes came with 2.0.

One harness, four ways in

Before, you picked an editor plugin or a terminal tool and lived with it. Now the same agent harness sits behind all of them. The IDE handles inline work: tab completion, targeted edits, diffs you accept or reject. The 2.0 app is where you leave long jobs running and check on them later. agy is a single Go binary for the terminal and for scripts. The SDK lets you build the same kind of agent into your own pipeline.

Subagents and worktrees

The thing that used to kill long agent sessions was the context filling up. The agent forgets the requirement from an hour ago and starts editing the wrong file. Subagents are the fix. Each one gets its own context window, does its piece and hands back a result. In the SDK you either let the main agent spawn them on its own or define them up front with their own instructions and tools. Restrictions on the parent carry down to the children, so a subagent cannot run terminal commands if its parent is not allowed to.

Projects can also start in a fresh Git worktree, so an agent trying a risky refactor does it on its own copy and leaves your working tree alone.

MCP, skills, rules and hooks

Tools used to be hardcoded. Now Antigravity plugs into MCP servers like everyone else, and teams can keep skills, rules and lifecycle hooks in the repo next to the code. Google also shipped Science Skills on the same day as 2.0: an open source bundle that ties the agent to more than 30 life science databases and tools, including the AlphaFold Database, the AlphaGenome API, UniProt and InterPro.

How it shows you things

This is where it feels most different from a chat window. Results come back as artifacts you can read and comment on. Sidecar links show up as small pills and open in a side pane, so terminal noise and background agent status stay out of the main conversation. Diffs are shown over the code.

Safety

It is no longer a choice between asking about everything and allowing everything. There is a terminal sandbox, a permissions system, and an allowlist and denylist for what the built-in browser may visit.

02What pushed it

Better Gemini models. Long context goes back to Gemini 1.5, which handled a million tokens. Since then Google has kept shipping: Gemini 3 at launch, 3.5 Flash as the 2.0 default, 3.8 Flash in September 2026, and Gemini 4 Argon reaching general availability on 13 October 2026. Each step made it cheaper and more realistic to let an agent read a whole codebase and keep going for longer.

Competition. Cursor, Claude Code, Devin and Windsurf changed what developers expect. Nobody wants a sidebar that hands you code to paste anymore. People want the tool to run the tests, read the error and fix the file.

Companies wanting guardrails. A business will not hand an agent a terminal that can delete files, leak keys or call out to the internet. Sandboxing, permissions and hooks are what make it sellable to them.

Google's own stack. Google wants the agent to work with Google Cloud and with DeepMind's science work. The Science Skills launch shows that clearly. A plugin system lets domain people add that without retraining the model.

03Look and feel, and how creative each one is

This section is opinion.

How they look

Antigravity 2.0 is clean and dense. What I like most is the side pane: logs, background agents and scheduled runs live there, and the main conversation stays readable.

Cursor is pure VS Code, dark and compact, with almost nothing on screen except your code. It is fast and stays out of the way. It does not try to show you rich visual output.

Devin Desktop, formerly Windsurf, was known for the Cascade timeline that stitched searches, commands and diffs into one stream. Cognition renamed Windsurf to Devin Desktop on 2 June 2026 and retired Cascade on 1 July 2026 in favour of a new local agent called Devin Local. It now feels more like an agent manager wrapped in an editor.

Claude Code started as a terminal tool and still feels best there: plain, quiet, no clutter. It has grown past that though. It now runs in a desktop app, on the web and in IDEs, and the desktop app has a built-in browser pane and can publish rich pages.

Devin in the cloud is a web dashboard where you watch the agent work in its own remote machine. You see everything, but it feels far from your own laptop.

How creative they are

Antigravity is strongest when you ask it to design something from nothing: a full prototype, a data pipeline, a diagram that explains a system. Claude Code is strongest at careful reasoning about code: clean refactors, sensible tradeoffs, small elegant fixes. Cursor is about guessing your next move, more muscle memory than design partner. Devin is creative in a stubborn way: drop it in an unfamiliar setup and it works around missing pieces until something passes.

04Where Antigravity sits

Copilot Workspace, which the original draft listed, ended its preview on 30 May 2025. Its ideas carried into the Copilot coding agent.

05Better at, and not better at

06Where this leaves things

The past year moved the whole field from chatting about code to running agents that do the work. Antigravity moved further than most. It went from one IDE in November to an app, a terminal agent, an SDK and editor extensions by May, all on the same harness, with subagents, scheduling, sandboxing and a plugin system on top.

Cursor still wins on fast, tight editing. A cloud agent like Devin still wins on keeping the agent away from your machine. Antigravity's place is the big, messy, multi-step job where you want several agents working at once, output you can actually read, and hooks into Google Cloud or science data. For that kind of work it is one of the strongest options right now.