Google Tests Agent Swarm Mode in Antigravity: Multiple AIs Write Code Together — But Who Cleans Up the Mess?

As reported by Alexey Shabanov of TestingCatalog, Google is testing a mode called Agent Swarm for Antigravity. Original article link: TestingCatalog.

Antigravity is Google's agent-first development environment, focused on letting AI agents run tasks across the code editor, terminal and browser, while leaving verifiable, auditable artifacts. It already caught developers' attention before, so the news of this new Swarm mode is well worth a look.

The term "Swarm" is not new. When applied to programming tools, its meaning is straightforward: let multiple agents work at the same time, or split work to collaborate. One agent is responsible for breaking down tasks, several modify code in parallel, another runs tests, and all results are aggregated in the end. It sounds just like porting the commonly seen multi-agent orchestration frameworks directly into the IDE.

Confirmed information is very limited for now. The report comes from spotted test traces, and Google has not released official feature specifications. Do not treat it as a publicly launched capability yet. What we can confirm is that Antigravity's product direction has long been focused on agent management. If the Swarm mode goes live, it will extend this management scope from a single individual agent to a whole group of agents.

Its actual impact on developers depends on three key details: how agents share context with each other, how conflicts are handled, and who is responsible for accepting the final output. It is not hard to get multiple agents to write code in parallel, the hard part is making sure they do not step on each other's toes. A codebase is not a chat history: file locks, dependencies and test failures will all amplify collaboration costs. If Antigravity only adds a single swarm toggle switch, its value will be very limited; it will only be worth trying if it can clearly show each agent's changes, assigned task boundaries and exact causes of failure.

Regular users can also understand it this way: before you hired one AI assistant, now you might hire an entire small team. Teams are more efficient, but they also come with higher management costs. Google has not released any timeline or pricing information. The original TestingCatalog article only provides early clues, not an official product launch announcement. We will wait and see the demos and usage restrictions Google rolls out next.