Google introduced Antigravity as an agentic development platform Gemini 3 as a top layer. The IDE isn’t just an autocomplete but also an environment in which agents execute complex software tasks and plan them across the editor, browser, and terminal surfaces. Google launched Antigravity on 18 November 2025 along with Gemini 3 to push towards agent-centric tools.
Antigravity – What Is It?
Google has described Antigravity, a new platform that combines an AI-powered IDE with a new approach to agentic development. It is the goal to transform the IDE into an agent-centric future. This will include browser control as well as asynchronous interactivity patterns, which allow agents to autonomously plan software tasks.
Antigravity behaves and looks like an AI editor, but it treats the agents as workers of first-class. Agents are able to break tasks, work with others, edit documents, execute commands and run a web browser. Developers operate at the task level while the system handles the tool interaction on the lower level.
Antigravity, under the hood is an Electron app based on Visual Studio Code. The application is available for free as a preview and requires that you sign into your Google account.
Models, Pricing and Runtime Environment
Antigravity allows agents to use multiple foundation models within the same framework. In the current version, agents are able to use models such as Gemini 3, Anthropic Claude Sonnet 4.5 or OpenAI GPT OSS. The developers can choose from a variety of models within one IDE, rather than being tied to a specific vendor.
Antigravity for individual users is free. Google says that the Gemini 3 Pro is subject to generous usage limits, which are refreshed every five hours. It also notes that only small numbers of users will be able to reach them.
Editor View And Manager’s View
Two main working modes are available to suit different neuronal models. These are consistently referred to as the Editor view and the Manager view in documentation and media coverage.
By default, the Editor view will be displayed. The interface looks similar to a normal IDE, with the agent on the side. Agents can edit and read files and suggest changes directly in the IDE. They also have access to terminals and web browsers when necessary.
This view allows for multiple agents to be viewed at once. It is here that you can coordinate the running of multiple agents rather than editing each line.
Artifacts, Not Raw Tool Logs
Artifacts are a crucial design feature in Antigravity. In place of showing only the raw tool logs, Antigravity agents create artifacts that are readable by humans and summarize what they’re doing.
The artifacts can be categorized into task lists and implementation plans. Other examples include screenshots, browser recordings, or walkthrough documents. These artifacts represent the work on a task-level rather than an API level, and they are intended to make it easier for developers verifying than dense traces.
Google is claiming that this tool was created to address the trust issue in existing agent frameworks. Some tools are either too detailed and overwhelm users or they hide all the steps to only display the final diff. Antigravity attempts to be in between by showing task-level artifacts and enough verification signals that allow a developer to audit what an agent did.
Four design tenets and feedback channels
Antigravity was built on four core principles: autonomy, self-improvement, feedback and trust.
The trust is managed through the artifacts, and steps of verification. The agents are given access to various surfaces including editor, terminals, browsers, and other editors. This allows them to run complex workflows with less prompting. Artifacts can be commented on to provide feedback, while agents are encouraged to learn from their past experiences and reuse successful procedures.
Developers can comment on artifacts directly, such as text or screenshots. This feedback can be incorporated into the work of agents without affecting their current task. You can correct an error without having to restart the entire task.
Agents can also access a feature that allows them to save code snippets or steps taken from previous tasks. In time, the internal playbook becomes something that agents are able to refer back to, instead of having them rediscover strategies with each project.
The Key Take-Aways
- Antigravity, an agent-first development platform, transforms the IDE’s interface into a plane of control where agents can operate on editor, terminal, and browser surfaces instead of a small inline assistant.
- This system, a fork of Visual Studio Code, runs on Windows, macOS, and Linux as a public preview with generous Gemini 3 Pro rates limits, and the option to use Claude Sonnet 4.5 or GPT OSS.
- Antigravity offers 2 primary modes. The Editor view allows for direct coding and includes an agent sidebar. While the Manager view is used as a control panel to manage multiple agents, workspaces or missions.
- The Agent can emit artifacts such as task lists, plans for implementation, screenshots, recordings of the browser, and much more. These act as verifiable proof of work and allow asynchronous workflows.
- Developers can add Google Docs-style comments to any artifacts on the surface. Agents then incorporate feedback from this and use a knowledge base for development without having to restart tasks.
Google Antigravity represents a practical step towards agentic development. Gemini 3 Pro anchors itself within a real IDE’s workflow. It exposes an Editor and Manager view to supervising agents and ensures that tasks are visible through Artifacts. Trust, autonomy, feedback and self-improvement are based on verifiable outcomes, not opaque traces. Antigravity views the IDE more as a governed space for autonomous agents than a window that displays code.
Take a look at the FULL TECHNICAL DETAILS here. Check out our GitHub Page for Tutorials, Codes and Notebooks. Also, feel free to follow us on Twitter Don’t forget about our 100k+ ML SubReddit Subscribe Now our Newsletter. Wait! Are you using Telegram? now you can join us on telegram as well.

