Google lets a team of AI agents write code together in Antigravity

On Tuesday, Google made its AI model Gemini 4 Carbon available in Antigravity, Google's coding platform in which AI programs write code themselves. At the same time, a test version of a new feature arrived: /swarm. With it, not one AI agent works on a coding task, but a whole team that divides the work and checks each other's work.

How the team works

The conversation the user is already having with Antigravity becomes the coordinator: the agent in charge. It cuts the task into pieces that do not get in each other's way. It gives each team member a name and writes down, for each one, what it has to do, what it must not touch and when it is finished.

There are also checkers. According to Google's documentation, they read the changes to the code and run the tests: small programs that check whether the code does what it should. If they find a mistake, they report it to the team member that wrote the code, which then fixes it. Usually there is one checker for every two to four builders.

All team members post their tasks, questions and test results on one shared message board. A team member that has to wait goes to sleep and then uses no tokens, the small pieces of text by which AI use is charged. It wakes up automatically as soon as another team member addresses it. A separate watcher steps in only if all team members have gone quiet before the work is finished.

Watching along and choosing models

In the Antigravity app, the user follows the message board in a side panel, with a list of the team members. A diagram shows who has addressed whom, and the messages can be replayed. In the version operated with typed commands, the command line, the board can fill the whole screen.

In its announcement, Google writes that the builders each work in their own copy of the code, so that their changes do not clash. The documentation on the site is more careful: by default they work in the same folder, and the coordinator can give each one its own copy if they edit files at the same time.

Each team member can use a different AI model. Google uses the heavy Gemini 4 Carbon for leading and checking, and lighter, faster models for the ordinary writing work. According to Google, Carbon ranks first in two tests made by other companies: CursorBench 4.0 for programming and Perplexity's WANDR for search and research work. When Google presented Carbon at the end of October, only security experts got access at first, because the model is especially good at finding and fixing security holes in software. Google does warn that a team of agents uses up a subscription's allowance much faster than a normal conversation.

Who can use it

According to the company, Google engineers used /swarm themselves for five weeks, for example to rebuild large pieces of software and to track down the cause of difficult bugs. Google does not give figures on those results.

The test version of /swarm is now becoming available in the Antigravity app and the command-line version. Subscribers to the paid plan Google AI Ultra can use Gemini 4 Carbon there, and also in Antigravity's extensions for coding programs. Scientists and researchers can apply for a chance at six months of free access to Carbon and to Gemini 4 Argon, the model that came out two weeks earlier.

What does this mean for the road to the singularity?

AI programs have been writing code for some time, but usually one works on a job at a time. With /swarm they divide a large coding task among themselves and one agent checks the work of another, while a human can read along on the message board and steer. This lets AI take on larger software projects than one agent can handle clearly on its own. The first users will now show how well that works in practice.

What is the technological singularity? Read the explanation →