Google unveiled Antigravity today alongside the launch of Gemini 3 Pro — a new development environment built for what the company calls an “agent-first future.” Antigravity uses Gemini 3 Pro at its core but can also run third-party models, including Claude Sonnet 4.5 and OpenAI’s GPT-OSS.
Antigravity supports multiple AI agents, providing them with direct access to a code editor, terminal, and an integrated browser. This setup is designed to enable agents to work more autonomously across various tasks or projects.
A defining feature of the platform is its ability to document its own operations. As agents work, Antigravity generates Artifacts—items such as task lists, plans, screenshots, and browser recordings — that summarize what the agent has done and what it intends to do next. Google says these Artifacts offer a clearer, more verifiable record than raw logs of tool calls or action traces.
Two Views: Editor and Manager
Antigravity introduces two main UI modes:
- Editor View: The default experience, resembling modern IDEs like Cursor or GitHub Copilot. It provides a standard coding workspace with an agent panel on the side.
- Manager View: A new interface for orchestrating multiple agents simultaneously. Google likens it to a “mission control” dashboard where users can deploy, coordinate, and observe several agents working across parallel workspaces.
Feedback and Learning
Users can now comment directly on specific Artefacts, giving targeted feedback without interrupting an agent’s ongoing workflow. Google also says agents in Antigravity will be able to “learn from past work” — storing useful code snippets, procedures, or repeating task patterns to improve future performance.
In a demo, an agent built a simple flight tracker app, ran tests, and produced a browser-recorded evaluation of those results, illustrating how Artefacts support transparency.
Antigravity is launching in public preview on Windows, macOS, and Linux. It’s free to use and includes what Google calls “generous rate limits” for Gemini 3 Pro. Rate limits refresh every five hours, and the company expects only a “very small fraction of power users” to ever encounter them.
