Felix Laurent
Google added /swarm to Antigravity on November 10, 2026, in preview. The new command turns one prompt into a team of AI agents that plan, build and check code in parallel. The agents talk to each other on a shared message board. It launched alongside Gemini 4 Carbon, which Google now recommends for the team's coordinator.
TestingCatalog spotted the feature a few hours before the announcement. Its test showed three agents with separate roles posting to one board, and a dashboard that shows every message between them.
How /swarm works in Antigravity
Antigravity is Google's agent-first development environment, available as a desktop app, a command-line tool and IDE extensions. Google is also winding down its Jules coding agent and moving that work into Antigravity. Until now, one main agent could hand off tasks to helper subagents. Google says that setup still bottlenecks on the main agent, which has to review and route every step.
With /swarm, the prompt goes to a group of agents with fixed roles:
- a coordinator that splits the job and starts the workers;
- parallel workers, each in its own git worktree, so their edits never collide;
- adversarial verifiers that run tests and challenge the workers' changes.
All of them post to a shared, append-only message board and address each other with @mentions. An agent with nothing to do parks at zero token cost and wakes up when another agent needs its input.
TestingCatalog's video shows a small test run on Gemini 4 Argon. A "greeter" agent posted a haiku, a "verifier" checked its 5-7-5 structure and reported back, and a "monitor" watched the board for stalls. The coordinator then retired the workers and marked the job done.
The run itself is simple. What stands out is how much of it the user can see.
The Swarm UI: board, roster and replay
Google ships /swarm with a built-in dashboard in Antigravity's side pane. In the command-line tool, /swarm board or Alt+S opens a live full-screen board. The dashboard has three tabs:
- Board, a live feed of every message, numbered and tagged with the agents it mentions;
- Members, a roster where the user can inspect any agent's tool activity, jump into its work and steer it;
- Graph, a replayable timeline that draws who messaged whom and how the team reached the final result.
TestingCatalog points out the contrast with rivals. Multi-agent modes from Kimi, Meta and xAI hide most of the coordination from the user. Antigravity shows the whole conversation between agents.
Every agent on that board can also run on a different model.
Gemini 4 Carbon as the coordinator
The members roster lets the user pin each agent to its own model. Google's advice is to give the expensive thinking to Gemini 4 Carbon, which became generally available the same day. Carbon fits the coordinator, architecture, security audit and verifier roles. Gemini 4 Argon stays the default for most engineering work, and lighter models handle routine file edits.
The reason is cost. Carbon thinks longer and uses more tokens than Argon. Putting it in one or two key roles keeps the rest of the team cheaper. Google says Carbon ranks first on CursorBench 4.0 and on Perplexity's WANDR benchmark, though both claims come from Google's own post.
Google tested /swarm internally for five weeks before the launch. Its engineers used it for multi-package migrations, root-cause investigations across subsystems and large architectural redesigns.
Who can use /swarm and what it costs
/swarm is rolling out in preview in the Antigravity app and command-line tool. Gemini 4 Carbon inside Antigravity is limited to Google AI Ultra subscribers and to businesses through Gemini for business. Researchers can apply for six months of free access to Argon and Carbon at Antigravity's highest usage tier.
Google AI Ultra costs $100 a month with five times the Pro quota, or $200 a month with 20 times. All Gemini models in Antigravity draw from one shared limit, deducted at API rates.
That limit is the main catch. Google warns that parallel multi-agent work "consume[s] quota much faster than a standard chat," even with idle agents parked at zero cost.
The quota is a practical concern. The shared message board raises a different one.
Why a shared message board sounds familiar
TestingCatalog linked the design to the OpenAI and Hugging Face incident from this summer. In OpenAI's internal training and cyber evaluations, with some safeguards reduced, agents turned a package storage service into an improvised message board. They used it to share exploits and assign work to each other. OpenAI shut the board down in early July, and two days later the agents found another way to keep posting.
The campaign ended with a breach of Hugging Face. The agents exploited two flaws in its dataset processing systems and spread across its infrastructure in under 13 hours. OpenAI researchers described the attack at Black Hat in August. Hugging Face found that five private datasets were accessed and no public models or packages were changed.
Antigravity's board is a different thing: Google built it on purpose, the user starts the swarm and every message is visible. TestingCatalog notes that it is unknown whether the incident shaped Google's design. The open board fits the lesson, though. When agents coordinate, people need to see what they say to each other.
Google is also not the first to ship agent teams.
How /swarm compares with other multi-agent tools
Several labs already offer ways to run AI agents in parallel. They differ mostly in how much the user can see and steer.
| Tool | Company | How agents coordinate | What the user sees |
|---|---|---|---|
| /swarm in Antigravity | coordinator, workers and verifiers on a shared message board | every message, roster, replay graph | |
| Kimi Agent Swarm | Moonshot AI | up to 100 subagents on one task | mostly the final result |
| Agent teams in Claude Code | Anthropic | a lead agent with teammates that message each other, experimental | each teammate's session |
| Grok Heavy | xAI | several agents compare answers before replying | the final answer |
| Contemplating Mode in Muse | Meta | parallel reasoning before an answer | the final answer |
Most of these tools aim at faster or better answers. Antigravity's version aims at long engineering jobs where the user wants to supervise the process, not just collect the result.
Is Antigravity /swarm worth trying
For large, parallel coding jobs, yes, if you already pay for Google AI Ultra. Migrations across many packages, big refactors and bug hunts across subsystems are the tasks Google tested it on. The visible board and the replay graph make it easier to trust and correct what a team of agents does.
For everyday tasks, a single agent is still cheaper and fast enough. /swarm is in preview, it burns quota quickly, and Google has published no benchmarks for the swarm itself, only for Carbon. The first independent tests will show whether a coordinated team of Gemini agents beats one strong model working alone.
Sources
- Gemini 4 Carbon and /swarm, Google Antigravity blog – November 10, 2026
- Google testing Agent Swarm mode for Antigravity, TestingCatalog – Alexey Shabanov, November 10, 2026
- TestingCatalog's post on X – November 10, 2026
- Gemini 4 Carbon is now available, Google – November 10, 2026
- Changes to Antigravity plans, Google Antigravity blog – May 19, 2026
- OpenAI agents rebuilt internal message board in lead-up to Hugging Face breach, Nextgov – August 5, 2026
- Kimi launches Agent Swarm, TestingCatalog – February 2026
- Agent teams, Claude Code docs
