Google Antigravity Agents Build Functional Operating System with Gemini 3.5

Google Antigravity announced on its blog that its multi-agent system successfully built a functional operating system from scratch using Gemini 3.5 Flash. The project used 93 subagents, more than 15,000 model calls, and roughly 339 million input tokens, costing about $917 at API rates. The OS was capable of running the open-source game FreeDoom.

The experiment demonstrated the potential of autonomous agent teams working asynchronously without human supervision. Each agent within Antigravity 2.0 had a specialized role, including Orchestrator, Worker, Reviewer, and Auditor, enabling parallel collaboration and self-healing behavior. The system used built-in mechanisms to handle stalled processes, context limitations, and code validation.

Beyond the operating system, the agents also reproduced a simplified version of the AlphaZero reinforcement learning system, including a playable interface. The team later extended the approach to create a photo editor, messaging app, and collaboration tool, all functional though not production-grade.

The same orchestration used in these experiments is now available through a new command called /teamwork-preview. The feature is offered as a research preview for Antigravity users on the Google AI Ultra plan, with support for Gemini models such as Gemini 3.5 Flash. Users can run autonomous agent teams locally, provided their machines remain active throughout the process.