Google is testing an Agent Swarm mode inside Antigravity

Coding with an AI assistant used to feel like managing an over-eager intern: you type a prompt, watch it generate a snippet, and constantly nudge it in the right direction.

Google is now testing a setup where you hand the project over to an entire team instead. A new Agent Swarm mode is in development for Google's Antigravity platform. Triggered by typing /swarm, the feature opens a dedicated dashboard that lets developers observe how multiple AI agents communicate, coordinate, and tackle work together.

Instead of keeping one conversation thread open for hours, the swarm uses a shared message board to distribute responsibilities across specialized roles.

Specialized builders and adversarial critics

Details around Antigravity's multi-agent system reveal a layered hierarchy split into orchestration, implementation, and verification.

Once a prompt is cleared, a coordinator named Sentinel takes charge and spins up a Project Orchestrator to break the request into modular milestones. To prevent context degradation on multi-day tasks, the orchestrator hands off duties to a fresh successor between phases.

Down in the trenches, read-only Explorer agents map out code repositories and call chains, while Worker agents write code and run terminal commands in isolated directories.

To keep the code reliable, Google built in adversarial verification gates. A Critic agent reviews code quality, while a Challenger agent actively writes edge cases and stress tests to break the solution. An Auditor checks real execution output to make sure tests actually ran rather than being faked or skipped, before a Success Auditor runs final checks.

Deep work, real demos, and heavy quota burn

The swarm capabilities tie into Google's latest Gemini 4 models. While Gemini 4 Argon handles everyday knowledge work, the newly released Gemini 4 Carbon is built for complex tasks like multi-package migrations and mathematical theorem proving.

Early internal trials highlight the speed gains. In one test, an employee used the multi-agent workflow to translate more than 100 web pages into 17 languages in just two days—a project that usually requires weeks of manual translation. Another demo used builder and critic agents working together to generate a photorealistic 3D temple inside Blender from a single prompt.

However, having several autonomous agents talking and arguing in parallel comes at a steep price. The interface specifically warns users about rapid quota consumption. Gemini 4 Carbon currently costs $2 per million input tokens and $10 per million output tokens, and Google has already scheduled those rates to double after November 30.

Why it matters

For engineers, this shifts human work from hands-on coding to high-level evaluation. You spend less time correcting syntax and more time setting boundary conditions and reviewing milestone gates. For trivial tasks it is massive overkill, but for massive refactors or multi-language rollouts, it offers genuine leverage.

Across the industry, multi-agent frameworks have mostly existed as flaky open-source experiments prone to infinite loops. Google integrating an orchestration board, role segregation, and automated critics directly into a development platform sets a higher bar for production agent workflows.

We think pitting Challenger and Auditor agents against the Workers is a clever way to catch hallucinated code and fake test passes. That said, keeping dozens of parallel agents aligned without context drift remains tricky. With API pricing set to jump soon, running a swarm will only make financial sense if it consistently produces production-ready results without human rescue.