Google unveils Antigravity, AI-driven coding platform with Gemini 3

Google has introduced Antigravity, a next-generation development platform designed to let AI agents operate like full-fledged software engineers. Built on the capabilities of Gemini 3 Pro and additional model integrations, the system allows AI to directly manipulate code editors, terminals, and web browsers, giving agents the ability to independently create, test, and validate software.

The release marks a major shift in how AI collaborates with developers. Instead of acting solely as assistants, agents in Antigravity can follow end-to-end workflows, from writing code to debugging and verifying results.

Gemini 3 Pro serves as the platform’s core engine, boasting a 1,501 Elo score on the LMArena benchmark and 76.2 per cent on SWE-bench Verified. Antigravity also supports Claude Sonnet 4.5 and GPT-OSS models, and is available across Windows, macOS, and Linux systems.

A key new feature in Antigravity is Google’s “Artifacts”, compact summaries of an agent’s activity such as task outlines, screenshots, and browser recordings. These replace long, unreadable action logs with clear, reviewable documentation that makes it easier for developers to understand and audit what agents are doing.

TWO WORK MODES TAILORED TO DEVELOPER NEEDS

The platform provides two distinct interfaces. Editor view mimics a standard IDE with an AI agent integrated into the sidebar, while Manager view functions like a mission-control dashboard, enabling users to oversee multiple agents working in parallel across different projects. Throughout the process, Antigravity captures its own visual history, screenshots and recordings, which developers can annotate with comments similar to Google Docs, letting them redirect an agent on the fly without halting progress.

Antigravity is built on Microsoft’s VS Code framework, meaning it supports thousands of existing extensions while layering on new capabilities designed specifically for agent-driven work. A dedicated Chrome extension lets agents execute and inspect code in a real browser environment, particularly useful for web applications requiring visual checks and UI validation.

The platform is guided by four core principles: transparency through clear reporting, autonomy via direct tool access, continuous feedback through inline comments, and self-improvement as agents learn from earlier tasks. Google says Gemini 3 Pro requests come with generous usage limits that reset every five hours, noting that only a small percentage of heavy users are likely to reach the cap.