Google Antigravity

What Antigravity is

Google Antigravity is an "agent-first" software development platform that Google released as a free public preview on 18 November 2025, for macOS, Windows and Linux^[1][2]. It appeared the same day as Gemini 3^[1][2]. "Agent-first" means the tool is organised around AI agents that plan and carry out work — writing code, running commands, checking results — rather than around a person typing every line, with the developer supervising what the agents do.

Where it came from

The project has a visible commercial history behind it. In July 2025 Google hired Windsurf CEO Varun Mohan, co-founder Douglas Chen and several researchers, and paid $2.4 billion to license Windsurf's technology; Mohan joined Google DeepMind^[2]. Google did not say what Antigravity was built on. Developers who examined the product concluded it was derived from VS Code, and found references to Windsurf's Cascade agent system^[2].

At launch

The first release offered Gemini 3 Pro, Gemini 2.5 Computer Use and Nano Banana, and also Anthropic's Claude Sonnet 4.5 and OpenAI's open-weight GPT-OSS^[3][2]. Carrying models from two competitors inside a Google product is the unusual part: the platform is positioned as a place to run agents, not only as a showcase for one vendor's models. Computer Use models operate software through the same screens and controls a person would use, which is what lets an agent drive a development environment rather than only produce text.

Security

After launch, researchers reported prompt-injection and backdoor weaknesses in the platform^[2]. Prompt injection is an attack in which text an agent reads — a file, a web page, a code comment — is crafted to make the agent follow instructions its user never gave. In December 2025 Google added an optional Secure Mode that requires human approval for agent actions^[2]. Because it is optional, the trade-off between speed and oversight stays a choice for the user.

Antigravity 2.0

At Google I/O on 19 May 2026 Google released Antigravity 2.0, a standalone desktop app for running multiple agents in parallel, together with an Antigravity CLI and an SDK^[2][6]. In this version Gemini 3.5 Flash inside Antigravity could deploy collaborative subagents^[2][6]: one agent can break a job into parts and hand them to others working at the same time, which is how the parallel-agent design is actually used.

Beyond the desktop app

Antigravity also became something other software could call. It was offered as a managed agent in the Gemini API, as antigravity-preview-05-2026 from 19 May 2026 and then as antigravity-preview-09-2026 from 17 September 2026^[5]. On 13 October 2026 it was built into the Gemini macOS app, so users could start a coding agent that builds and tests code in the background^[4].

Where it sits

In shape, Antigravity went in under a year from a preview product launched alongside Gemini 3 to a set of entry points: a desktop app, a command-line tool, an SDK, a managed agent inside the Gemini API, and a built-in capability of the Gemini macOS app^[2][5][4]. The direction is to treat "letting agents write code" as a service that can be called and orchestrated, not only as a feature inside an editor window. The security discussion around it shows the central tension of such tools: the more permission and autonomy an agent gets, the more the product depends on approval and oversight mechanisms^[2].