Google Antigravity: Agentic Development Platform Release

Google DeepMind has announced Google Antigravity, an agentic development platform designed to transition software engineering from synchronous AI assistance to autonomous, agent-first workflows. By integrating browser control, asynchronous interaction patterns, and a specialized product form factor, Antigravity allows AI agents to plan and execute end-to-end software tasks with minimal human intervention.

Agent-First Architecture and Surfaces

Antigravity shifts the paradigm from embedding AI within an IDE to embedding development surfaces within the AI agent. The platform provides two primary interfaces to support different modes of interaction:

Editor View: A traditional AI-powered IDE experience featuring Tab completions, in-line commands, and a side-panel agent for synchronous work.
Manager Surface: A "mission control" designed for asynchronous orchestration. It allows users to spawn, observe, and manage multiple agents across different workspaces in parallel, enabling tasks like background research to occur while the user focuses on primary development.

Core Tenets of Collaborative Development

Antigravity is built on four foundational pillars to ensure that autonomous agents remain useful and controllable:

Trust through Task-Level Abstraction

To avoid the extremes of overwhelming users with raw tool calls or providing opaque final results, Antigravity uses task-level abstractions. Agents communicate progress via Artifacts—tangible deliverables such as implementation plans, task lists, walkthroughs, screenshots, and browser recordings. This approach emphasizes the verification of work, allowing users to validate the agent's logic and results before and after implementation.

Autonomy Across Surfaces

Leveraging models like Gemini 3, Antigravity agents can operate autonomously across multiple surfaces simultaneously. An agent can write frontend code, use the terminal to launch a localhost server, and actuate a browser to test the feature's functionality without requiring step-by-step user prompts.

Asynchronous Feedback Loops

Antigravity enables intuitive, asynchronous feedback to prevent the "80% completion" problem where the final 20% of a task becomes too costly to fix. Users can provide feedback directly on Artifacts—such as Google-doc-style comments on text or select-and-comment feedback on screenshots—which the agent incorporates into its execution without needing to stop the process.

Knowledge Management and Self-Improvement

Learning is treated as a core primitive. Agent actions both retrieve from and contribute to a shared knowledge base. This allows agents to learn from past work, storing explicit information like code snippets and architecture decisions, as well as abstract patterns, such as the sequence of steps required to successfully complete specific subtasks.

Availability and Model Support

Google Antigravity is currently available in public preview at no charge for individuals. It is compatible with MacOS, Linux, and Windows.

To provide developers with model optionality, the platform supports several frontier models:

Gemini 3 (including Gemini 3 Pro with generous rate limits)
Claude Sonnet 4.5 (Anthropic)
GPT-OSS (OpenAI)

Access to these models is subject to capacity-based rate limits that refresh every five hours, with limits correlated to the volume of work performed by the agent rather than a simple prompt count.