Google Antigravity /swarm Guide: Multi-Agent Teams, Remote Agents and Voice Mode (November 2026)

Google Antigravity can now turn one prompt into a whole team of AI agents. The new /swarm command arrived in version 2.31, and Gemini 4 Carbon landed a day later to lead those teams.

The same week also brought remote agents, voice mode, a Knowledge Work layout and download approvals for the browser agent. This guide shows how to set up and use each one.

What Is New in Antigravity This Week

Google shipped three Antigravity releases in about a week, then a new flagship model. Here is the order things arrived, based on the official changelog and Google’s launch post.

If you want the model side in more depth, my Gemini 4 Carbon benchmarks and access guide covers scores and plans. This post stays focused on using the features.

What /swarm Is and How the Team Works

A normal Antigravity chat is one agent doing one thing at a time. With /swarm, your current conversation becomes a coordinator. It splits your task, creates helper agents, and reports back to you.

Google’s launch post describes it as turning one prompt into a coordinated team of agents you can track in real time. The changelog adds that the agents coordinate over a message board for long-running tasks.

The four roles in a swarm

• Coordinator: splits the task, names the swarm, adds workers and reports to you. It does not do the technical work itself.
• Workers: named agents such as @parser, or numbered groups such as @fix-1 to @fix-4. Each gets a clean context and a brief that says what it owns, what it must not touch, and what “done” means.
• Verifier: usually one for every two to four workers. It reads diffs, runs tests and sends problems back to the worker who wrote the code.
• Monitor: stays asleep while work moves forward. It only posts if every member goes quiet before the job is finished.

The shared board

All members talk on one shared board. Assignments, questions, diff summaries and test results are posted there with @name, @group or @all mentions. Workers do not send each other private messages.

Idle members sleep and use no tokens until someone mentions them. In the App, the board has three tabs: Board, Members and Graph, where Graph can replay how the work flowed.

Note: Workers share your workspace by default. The coordinator can give workers their own copy of the repository, but removing that worker deletes its copy. Make sure its changes land first.

Who Can Use /swarm and Gemini 4 Carbon

The plan rules are different for the command and for the model. That trips people up, so here they are side by side.

So a Pro user can run a swarm, but the workers will use the other models your plan can pick. To put Carbon in the coordinator or verifier seat, you need Ultra. My Google AI Pro vs Ultra guide compares the plans.

What Carbon costs on the API

If you reach Carbon through the paid Gemini API instead, Google lists an introductory price of $2 per million input tokens and $10 per million output tokens. After November 30, 2026, it rises to $4 input and $20 output.

Warning: Each swarm worker is its own agent, so a swarm burns quota much faster than one chat. Google also says Carbon uses more tokens than Argon because it explores problems more deeply. Start with small tasks.

How to Start a Swarm (App and CLI)

Update Antigravity first. /swarm needs v2.31.0 or later of the App. Then follow these steps.

1. Open a project. Open the repo or folder you want the team to work on. Commit any work you care about so you can roll back.
2. Pick the base model. Workers default to your conversation’s model, so choose it in the model picker before you start.
3. Type the command. In the App, type /swarm, press Tab or Enter, then describe the task. In the CLI, type /swarm, press Tab to complete it, then describe the task.
4. Write a clear goal. Say what “done” means, which tests must pass, and which folders are off limits. The coordinator turns this into worker briefs.
5. Open the board. In the App, click the Open the Swarm UI link in the reply. In the CLI, run /swarm board or press Alt+S.
6. Approve requests. In the CLI, Alt+J jumps to the next worker waiting for approval and Ctrl+K approves it.
7. Review the result. Ask the coordinator for a summary, then read the diffs and run the tests yourself before you merge.

A good first prompt

Here is an example shape that gives the coordinator what it needs. Adjust the paths and tests for your own project.

/swarm Add input validation to every route in src/api.
Done means: all routes reject bad input with a 400,
npm test passes, and no files outside src/api change.
Use a separate verifier on Gemini 4 Carbon.

Tip: On macOS, Alt+S only works if your terminal sends Option as Alt. The /swarm board command works in every terminal.

CLI board keys worth knowing

• Move: j and k, or the arrow keys. g and G jump to the first and last item.
• Read: Enter shows details and Space expands a post.
• Filter: f follows new posts and p filters by member.
• Open: o opens a member’s conversation.
• Close: q or Esc. Alt+X hides the status row.

Steering, Models and Stopping Workers

You stay in charge the whole time. There are three ways to change course while a swarm runs.

• Talk to the coordinator: give it new scope or priorities and it posts instructions to the right workers. You can also ask it for status or blockers at any time.
• Message one worker: in the App, open it from the Members tab or Subagents list, or start your prompt with @name. In the CLI, run /agents to list and open workers.
• Change a model: ask the coordinator to move a waiting worker to another model. It keeps its conversation history.

Member states

Every member shows one of three states. Active means it is running a turn, Waiting means it is asleep on the board, and Killed means it was removed.

Killed members cannot come back. If you need that role again, the coordinator adds a new worker instead.

Stopping

The docs do not list one command that stops a whole swarm. Instead, ask the coordinator to remove workers that have nothing left to do. Waiting workers cost nothing, so you do not need to stop them just to save tokens.

/swarm vs /teamwork

Antigravity already had a /teamwork command for agent teams, and it is still documented. The two are built for different jobs.

If you use Claude Code too, the idea is close to running several subagents at once. My Claude Code subagents guide shows how that side works.

How to Set Up Remote Agents

Remote agents run your conversation in an isolated, Google-managed environment instead of on your computer. Tasks keep going when you close the browser, sleep the laptop or switch devices.

They need a personal Google account on AI Pro or AI Ultra. They are not offered on the free plan, Google AI Plus, or Workspace and Cloud business accounts.

Create a remote project in the App

1. Start a new remote project. In the left sidebar, click + next to Projects and choose New Remote Project.
2. Add resources. Click Add Resource to attach GitHub repos, Google Drive files or uploads. Then click Next, or Skip.
3. Name it. Type a name under Set Project Name and click Create.
4. Send a prompt. Select the project and give the agent its task.
5. Check from your phone. Open antigravity.google.com in a mobile browser and use the switcher in the top bar to find Remote Conversations.

Start one from the CLI

agy update
agy --remote-agent

You need CLI 1.4.12 or later. The first session creates a project called “CLI Project” with your machine name, and later sessions reuse it. Use /projects to switch projects during a session.

Limits to know

• No local files: remote agents cannot read your computer. Attach files to the project instead.
• Upload caps: 8 MiB per file, 50 files and 100 MiB per project, and 2 GiB per account.
• Modes: remote sessions always use the default mode. Plan, accept-edits and auto modes are not available.
• Permissions: the default preset is always-proceed (turbo), so commands run without asking. Change it with /permissions if you want reviews.
• Missing CLI features: /voice, /diff, /rewind, local MCP servers, plugins and skills do not work in remote CLI sessions.

How to Use Voice Mode

Voice mode is a two-way spoken chat with the agent, powered by Gemini Live. You can give tasks, ask for progress, or stop work just by saying so.

1. Clear the message box. The voice button only shows when the box is empty.
2. Click Start Voice mode. It is the round button next to the dictation mic. Allow microphone access if asked.
3. Pick a voice. The first time, a dialog asks you to choose one. A chime plays when it is ready.
4. Talk. Say “Wait” to pause the agent, or ask it to show a table on screen.
5. End the call. Click End, or say “Bye”. Running tasks keep going.

In the CLI (version 1.5.0 or later), run /voice live or press F6. Press F7 or run /voice stop to end.

Note: You cannot approve destructive actions such as rm, sudo or force-pushes by voice. You must click Allow on screen. You can still deny them by voice.

Voice mode works in one conversation at a time, and only in local conversations in the App. Since v2.31, a finished call shows a short summary on its transcript card.

Knowledge Work Layout and Other v2.30/2.31 Changes

The Knowledge Work layout is a simpler interface for research, writing and documents. It runs the same agent with the same models, plugins, skills and MCP servers. Only the screen changes.

How to switch layouts

1. Open Settings. Press Cmd+, on macOS or Ctrl+, on Windows and Linux.
2. Go to the layout option. Choose General, then Interface Layout.
3. Pick a layout. Choose Knowledge Work or Software Engineering. The change applies at once, with no restart.

In Knowledge Work, Git branch pickers and IDE buttons are hidden. The side panel focuses on Documents, Artifacts, Links, Uploads and Background Tasks, and permission prompts use plain language.

The rest of v2.30 and v2.31

• Browser download approval: when the browser agent tries to download a file, the download pauses until you approve it.
• No input lock: pages the browser agent controls can no longer lock your mouse pointer or keyboard.
• Font size: set 12, 14 or 16 px under Settings > Appearance.
• Split view: plugin views can sit beside a conversation.
• Role setup: on first sign-in, Antigravity asks what kind of work you do and sets a matching layout and starter prompts.
• Imports: bring MCP servers, skills, commands, subagents and instruction files from Claude Code, Codex and Cursor into a plugin you own.

Comparing Antigravity with other agent tools? See my Claude Code vs Gemini CLI vs Antigravity CLI comparison.

My Take

/swarm is the most interesting part of this update. The verifier ratio and the “no private messages” board rule show Google is trying to keep many agents honest, not just fast.

The catch is cost. Every worker spends its own quota, and Carbon is a heavy thinker. For most people, a small swarm on Argon with one Carbon verifier is the sensible place to start.

Remote agents and voice mode are quieter wins. Being able to start a repo task, close the laptop, and check it from a phone is the kind of thing you use daily.

Troubleshooting: Common Problems and Fixes

• /swarm does not autocomplete: update the App to v2.31.0 or later and check you are on AI Pro or Ultra. It is not on the free plan.
• Gemini 4 Carbon is missing from the model picker: Carbon is for AI Ultra in Antigravity and is still rolling out. Pro users can run swarms on other models.
• A worker stopped and never came back: it probably hit its quota. Paused members do not restart when quota refreshes. Post to it on the board, switch its model, or message it directly.
• Alt+S does nothing on a Mac: set your terminal to send Option as Alt, or just run /swarm board.
• Remote agents option is missing: you need a personal account on paid Pro or Ultra. Workspace accounts and AI Plus are not eligible. In the CLI, run /help to confirm your account.
• “is a local conversation” error in the CLI: you tried to resume a local chat with --remote-agent. Start the CLI without that flag.
• Start Voice mode button is missing: clear the message box, sign in, and check you have not hit your usage limit. It does not show in remote or cloud conversations.
• A download from the browser agent is stuck: that is the new approval step in v2.31. Approve the download to let it finish.

Frequently Asked Questions

Is Antigravity /swarm free?

No. /swarm is a preview feature for Google AI Pro and Google AI Ultra subscribers in the Antigravity App and CLI.

Do I need Gemini 4 Carbon to use /swarm?

No. Workers default to your conversation’s model, so you can run a swarm on any model you can pick. Carbon is suggested for the coordinator and verifier roles, and individuals need AI Ultra to use it in Antigravity.

How many agents does a swarm create?

The coordinator decides based on your task. The docs say there is usually one verifier for every two to four workers, plus a sleeping monitor.

Does a swarm use a lot of quota?

Yes. Each worker is its own agent, so a swarm uses quota much faster than one chat. Idle members sleep and use no tokens until mentioned.

What is the difference between remote agents and remote control?

Remote agents run in a Google-managed cloud environment, so nothing on your side needs to stay on. Remote control runs on your own desktop or server, which must stay on, but can use local files and tools.

Can I use voice mode on a work account?

Not yet. Since v2.30, voice mode is available to everyone except business and enterprise accounts.

How much does Gemini 4 Carbon cost on the API?

Google lists $2 per million input tokens and $10 per million output tokens until November 30, 2026. After that, it is $4 input and $20 output per million tokens.

Sources

• Google Antigravity changelog (v2.29 to v2.31)
• Google Antigravity docs: /swarm
• Google blog: Gemini 4 Carbon is now available
• Google Antigravity docs: Remote Agents
• Google Antigravity docs: Voice mode
• Google Antigravity blog: Antigravity for Knowledge Work
• Google Antigravity docs: Teamwork agent teams
• 9to5Google: Antigravity remote agents and voice mode