Google Antigravity is an agentic development platform published by Google, announced in November 2025. Google describes it as not just an editor but a development platform that combines a familiar AI-powered coding experience with an agent-first interface, from which agents can be deployed to autonomously plan, execute and verify complex tasks across the editor, terminal and browser.
The design premise Google states for it is that agents should have a dedicated space to work in rather than occupying a sidebar, and that a developer should be able to operate at a higher, task-oriented level rather than at the level of individual prompts and tool calls. Its answer to the resulting trust problem is to have agents report through reviewable deliverables rather than through raw tool-call logs. The product team’s own launch post (Introducing Google Antigravity, a New Era in AI-Assisted Software Development) organises the product around four tenets — trust, autonomy, feedback and self-improvement — and presents it as the IDE evolving toward an agent-first future, prompted by models such as Gemini 3 becoming able to run for longer periods without intervention across multiple surfaces.
The name has since come to cover a product family. The product changelog lists the original launch as release 1.11.2 on November 18, 2025, and records a second generation, Antigravity 2.0, launching in May 2026 with an import path from Antigravity 1.0; the changelog lists the Antigravity IDE as its own product line, whose 2.0.1 entry of the same month is described as the first release of the Antigravity IDE, and the family also includes the Antigravity CLI and the Python Antigravity SDK. The overview and capabilities below describe the platform as launched unless a version is named.
Capabilities
The platform presents two distinct surfaces. The Editor View is an AI-powered IDE with tab completions and inline commands, for the synchronous hands-on workflow Google describes as already familiar; the launch post adds that it carries a fully functioning agent in the side panel. The Manager surface is the agent-first interface, where several agents can be spawned, orchestrated and observed working asynchronously across different workspaces — Google positions it for dispatching long-running maintenance work or bug fixes in the background, including tasks that span reproducing an issue, generating a test case and implementing a fix. The launch post describes it as flipping the paradigm from agents embedded within surfaces to surfaces embedded into the agent, likening it to a mission control, and says the team deliberately kept the Manager and the Editor in separate windows, optimising for instant handoffs between them rather than squeezing both into one.
Agents work across the editor, terminal and browser within one task. Google’s worked example has an agent write code for a feature, use the terminal to launch the application, then use the browser to test and verify the new component, without synchronous human intervention.
The product documentation for the Antigravity IDE describes the browser side of this: the agent can open, read and actuate a local Chrome browser — to test development websites, read documentation or automate browser tasks — working through a specialised browser subagent that captures screenshots and saves action videos as artifacts. Browser tools can be switched off entirely in settings; URL access is controlled by a two-layer denylist and allowlist, and the agent runs in a separate Chrome profile from the user’s own.
Verification is mediated by Artifacts — deliverables the agent generates, such as task lists, implementation plans, walkthroughs, screenshots and browser recordings. Google’s stated rationale is that these let a developer check the agent’s logic at a glance where scrolling raw tool calls would be tedious; the launch post frames this as presenting agent work at a task-level abstraction, with tool calls grouped within tasks. Feedback can be left directly on an Artifact, comparably to commenting on a document — the launch post names Google-doc-style comments on text Artifacts and select-and-comment feedback on screenshots — and the agent incorporates that input without halting its execution flow. Google also states the platform treats learning as a core primitive, allowing agents to save useful context and code snippets to a knowledge base to improve future tasks; the launch post says that knowledge can also be more abstract, such as the series of steps that completed a particular subtask.
Later releases
Among the releases after launch, the changelog records agent Skills arriving in January 2026 (1.14.2), terminal commands
running inside a sandbox — first on macOS (1.15.6), later on Linux (1.21.6) — with file and
network sandboxing on Windows following in the 2.0 line (2.15.1), rules read from AGENTS.md as well as GEMINI.md (1.20.5), and a
unified agent permissions system (1.22.2, April 2026).
In the Antigravity 2.0 line, the changelog records Remote Control for driving local agent
sessions from any browser (2.9.1), an embedded terminal and Git version control in the
sidebar (2.10.0), inline “generative UI” rendering of HTML artifacts (2.11.0), a /boost
command for a multi-agent reasoning pass (2.12.0), and a /plan command whose plans can be
read and edited before any code is written, governed by a Plan Review Policy setting
(2.17.0).
Antigravity 2.0
The product documentation describes how Antigravity 2.0 organises work. Agents work in Projects, where previously in the Agent Manager each agent was mapped to a single workspace folder. A project can span several folders, supports Git worktrees so agents can operate in isolated background folders, and carries its own scoped settings and permission grants — so a trusted project can be given a more permissive security preset than an untrusted folder. Quick one-off conversations can also run outside any project, in an isolated local scratch folder.
By default the documentation describes agents as asking for explicit permission before
running terminal commands outside the terminal sandbox on macOS and Linux (or before any
terminal command on Windows), and as able to read and write only within a project’s
folders unless the user broadens that setting. Other 2.0 features it lists include
scheduled tasks that send messages to agents on a timer, live voice transcription for
prompts and Artifact comments, JSON hooks that run local shell scripts at points in the
agent’s execution cycle, a reworked browser subagent invoked with /browser that
integrates with Chrome DevTools MCP, Remote Control of desktop agent sessions from any web
browser, an integrated terminal, and a Git-native version-control panel for reviewing
uncommitted, branch and agent-made diffs.
Google’s “Getting Started with Google Antigravity” codelab describes the same generation as a family of products for what it calls the agent-first era rather than an Agent Manager attached to an IDE. At its centre is Antigravity itself, a standalone application for macOS, Linux and Windows that the codelab calls a command center for managing multiple local agents in parallel and running scheduled tasks, and which, unlike its predecessor the Agent Manager, works independently of an IDE. Around it sit the Antigravity IDE — the original agentic IDE, which the codelab says remains available and recommended for developers — the Antigravity CLI, the Antigravity SDK, and Antigravity for IDEs, extensions for editors such as Visual Studio Code, JetBrains IDEs and Zed.
The codelab explains projects as a combination of folders that defines both the environment and the
scope of an agent. Each project inherits its permissions — artifact review, allowed tools, MCP server
configuration — from a global configuration, and can override them individually: a security preset
deciding whether terminal commands and file accesses are reviewed first, whether an implementation
plan must be approved before the agent executes it, allowed file paths and allowed or blocked URLs,
and which MCP tools the project may use. MCP servers, local or remote, are configured in
$HOME/.gemini/config/mcp_config.json, and the codelab describes pre-packaged configurations for
Google Cloud services that install with one click, with individual tools switchable on and off.
It presents Artifacts as Antigravity’s answer to a trust gap — where a developer previously had to
read code to check an agent’s claim to have fixed a bug, the agent now produces an artifact to show
it — and lists the main ones as task lists, implementation plans, walkthroughs and screenshots, with
code diffs reviewable and commentable alongside them though technically not artifacts. Skills are
described as loading through Progressive Disclosure: only a skill’s metadata is read
up front, and its full instructions are loaded when a request matches its description. Global skills
live in ~/.gemini/config/skills/ and are available across Antigravity, the Antigravity IDE and the
Antigravity CLI; project skills live in <project-root>/.agents/skills/.
Subagents
The documentation for Antigravity 2.0 and the Antigravity CLI describes delegation to
concurrent background subagents. A parent agent spawns one with an invoke_subagent tool,
giving it a role and an initial prompt; the subagent starts with a clean context rather
than the parent’s conversation history, and can inherit the parent’s workspace, get an
isolated Git worktree, or share directory storage. Three subagents come built in —
research for codebase exploration, browser for interactive browser testing, and self,
a clone of the calling agent — and custom ones are defined as Markdown files whose YAML
frontmatter sets their tools, model tier and command-execution policy and whose body is
their system prompt.
A subagent is running, idle (finished and waiting, re-awakened by a message) or killed.
Agents message each other by conversation ID and can read each other’s transcripts, and
nesting is capped at ten levels below the primary agent. Subagents inherit the parent’s
command allowlist, file-access scopes and sandbox settings, and a request needing user
approval bubbles up to the main interface. On top of this, the documentation describes two
multi-agent orchestrators: /boost, a three-tier reasoning hierarchy for hard problems
within an interactive session, and /teamwork-preview, which coordinates a team of
specialised agents on larger projects through milestone decomposition, parallel
implementation and independent verification.
Adoption & Ecosystem
At its announcement the platform was released in public preview, at no cost for individuals, and cross-platform across macOS, Windows and Linux. Google stated it offered model optionality: generous rate limits on Gemini 3 Pro, with full support for Anthropic’s Claude Sonnet 4.5 and OpenAI’s GPT-OSS. This availability and model list describe the platform as announced in November 2025.
Writing on launch day, Simon Willison described the platform at first glance as “yet another VS
Code fork Cursor clone” — a desktop application that signs in to a Google account and provides an
IDE for agentic coding against Gemini models — but said that on closer inspection it was a fair bit
more interesting. He picked out its three surfaces (an agent manager dashboard, a VS Code-style
editor and deep browser integration through a new Chrome extension, which he compared in role to
Playwright MCP in letting the agent test the web applications it builds) and its Artifacts, which he
described as Markdown documents created automatically as the agent works. His own first attempt,
adding Gemini 3 support to his llm-gemini plugin, worked at first and then stopped with an
“Agent execution terminated due to model provider overload” error; he said he would try again once
the initial launch jitters had been worked through.