New AI coding assistants appear almost weekly, and the claims tend to sound alike: faster code, automatic bug fixes, whole applications from a prompt. Most of them nonetheless work the same way. There is a chat window, a model, a long answer, and then the tool waits for the next instruction. That is useful, but it is assistance rather than engineering.
Google Antigravity, which Google introduced in November 2025 alongside its Gemini 3 model, takes a different approach. Google describes it as an agent-first development platform rather than another coding assistant: an environment in which multiple AI agents work in parallel across the editor, the terminal and a browser, with the developer supervising rather than dictating each step.
From a linear chat to a team of agents
In a conventional AI tool, everything moves in a straight line. The developer assigns a task, the model reasons, writes code and stops. Antigravity breaks that pattern by running several agents at once, each with its own task, state and pace. The experience, as Google frames it, is closer to managing a small engineering team inside the editor than to chatting with a single assistant.
Three surfaces
Antigravity is organized around three main surfaces. The first is the Agent Manager, which functions as a kind of mission control. It shows every running agent, its task, its progress and its reasoning, and lets the developer pause one task, leave another running in the background or launch new work, much as a lead would coordinate a team.
The second is the Editor. It looks like an ordinary code editor, but the agent maintains awareness of the whole codebase. Ask for a refactor, a migration or a bug fix and the agent does not simply reply with a suggestion; it opens files, modifies functions and completes partially written code. Google says agents produce what it calls artifacts, such as task lists, implementation plans, screenshots and browser recordings, so that the developer can verify what was done without reading every line.
The third, and the one Google emphasizes most, is browser control. Antigravity can drive a real browser: clicking, filling in forms, scrolling and recording the full flow. The instruction “test my feature” becomes an action rather than a request for a description. The agent opens the application, interacts with it and reports back, reducing the need for screenshots and long written context from the developer.
Levels of autonomy
Antigravity offers several modes that trade control for speed. The most ambitious is one in which the agent makes most decisions itself and asks for approval only at significant points. The developer sets strategy while the agents handle the detailed work. As with any autonomous coding tool, that mode depends on the developer reviewing the artifacts the agent produces, since errors made at speed can propagate across a codebase.
A multi-model approach
A distinguishing feature is model choice. According to Google’s launch announcement, Antigravity gives developers access to Gemini 3 Pro, Anthropic’s Claude Sonnet 4.5 and OpenAI’s open-weight GPT-OSS models within the agent, and the models can be assigned to different kinds of work. In practice, developers have paired one model for front-end and interface tasks with another for back-end architecture and complex logic, treating the models like specialists on a team. Google launched the product in a free public preview with generous rate limits, though the company has said usage limits apply and pricing may change after the preview period.
Whether this model of development holds up on large production codebases is still being tested by early users. But Antigravity illustrates where the major vendors are heading: away from a single assistant answering questions and toward environments in which the developer’s main job is to direct and review the work of several agents.
