Google Antigravity launched: cross-model free switching, AI agents autonomously executing multiple projects.

Google launched its new AI agent-specific integrated development environment (IDE) “Google Antigravity” simultaneously with the release of Gemini 3, and opened it for public testing. Antigravity is centered around Gemini 3 and supports models such as Anthropic Sonnet 4.5 and OpenAI gpt-oss, emphasizing an “agent-first architecture” (Agent-First), enabling AI to no longer be just a passive tool, but to autonomously perform complex tasks such as cross-file analysis, asynchronous processes, and result verification.

Agent directly controls the interface, shifting from a collaborative role to an autonomous execution role.

The core concept of Antigravity is to allow AI to no longer just be a tool that is called upon, but to be able to execute tasks itself within the interface. It comes with a built-in IDE-like editor (Editor View), where the AI agent is embedded directly, capable of browsing code, clicking interfaces, switching files, and viewing errors, making the overall operation almost identical to that of the engineer himself.

The Antigravity interface is similar to coding resource websites like GitHub Copilot, where the AI agent helps users write code.

Google also announced that the next step will be to launch the “Manager Surface” (Manager Surface), which will embed the interface back into the agent, further enhancing the agent's task management capabilities. The special aspect here is that:

“In the management interface, users can simultaneously schedule and control the project progress of multiple AI agents, operating like a mission control center.”

Support complete asynchronous processes, allowing project tasks to run automatically.

Antigravity provides a complete asynchronous coding workflow, allowing project tasks to automatically progress and return results even when human developers are offline and the IDE is not open. It can handle large review or refactoring processes across files and steps, making it particularly suitable for teams that require frequent code compilation.

At the same time, Antigravity also emphasizes the “Verification-First” ( Verification-First ) concept, where every time the AI agent completes a task, it will provide the necessary verification information, including a task summary, supporting documents, and testing or inspection procedures.

Google emphasizes that the platform will present the entire execution process with a “task-level abstraction,” avoiding black boxes and excessive information, allowing users to clearly understand the quality of the agent's work.

( Note: The task level is abstract, meaning that AI will not overwhelm you with all the granular execution steps at once, nor will it say nothing at all. Instead, it presents the necessary and just-enough information to the user in a task-based manner. )

Can freely switch between models and supports knowledge base enhancement for self-learning.

During the public beta, Antigravity supports three major models:

Gemini 3

Anthropic Sonnet 4.5

OpenAI gpt-oss

Developers can switch freely according to work needs. AI agents can also connect to the team's knowledge base, allowing them to retrieve past cases while working and write back new learned information to the knowledge base after completing tasks, gradually improving the performance of subsequent tasks.

This article Google Antigravity is online: cross-model free switching, AI agents autonomously executing multiple projects. First appeared in Chain News ABMedia.