01What Antigravity Swarm Mode Is and How It Works

Antigravity swarm mode is a new feature in Google Antigravity that turns one chat into a team of AI agents. You give it one prompt, and a coordinator splits the job and hands each part to its own worker agent.

Before this, one chat meant one agent doing one thing at a time. Now your chat becomes the coordinator. You type /swarm, then your task.

The coordinator breaks the job into parts, spawns parallel workers, and gives each worker its own isolated space so they do not get in each other's way.

In this episode, Julian Goldie used it to build what he calls a hive. One prompt started nine AI agents. Each one built a piece of a real business app, while two inspector agents checked every piece.

By the end of the episode, there were three working systems.

Swarm mode arrived together with Gemini 4 Carbon. Google describes Carbon as its strongest model for long software jobs. Swarm mode is still in preview, and you can use it in both the Antigravity app and the Antigravity CLI.

02Main Features of Antigravity Swarm Mode

Swarm mode is more than a way to run many agents. It also gives you a panel to watch and control them while they work.

  • Coordinator chat: your chat plans the job and briefs every worker.
  • Parallel workers: each worker builds its part at the same time as the others.
  • Isolation: every worker gets its own space, so their work does not clash.
  • Sleeping agents: when an agent has nothing to do, it sleeps, so it is not burning anything.
  • Verifier agents: inspectors check the finished work and send it back if it breaks.

On the right side of the screen you get the swarm pane. It has three tabs:

  • Board: the live feed of what every agent is doing.
  • Members: a list of every agent. You can jump in and steer one, or pin it to a model.
  • Graph: lets you scrub back through the whole run and see who sent what to whom, and when.

Together, these make a big parallel job easy to follow instead of a black box.

03How to Set Up Antigravity Swarm Mode

Getting started takes a few minutes. Antigravity runs on Mac, Windows and Linux. Here are the steps Julian walks through in this episode:

  1. Go to antigravity.google/download and install the app.
  2. Open it and click New Conversation.
  3. Add your project folder.
  4. In the model picker, choose Gemini 4 Carbon.
  5. Set permissions to auto, so it does not stop to ask on every file.
  6. Type /swarm, paste your prompt and press Enter.
  7. When the coordinator replies, click to open the swarm UI.
  8. Watch the Board and check the Members tab to see who is working.

One thing to know: Gemini 4 Carbon is rolling out to Google AI Ultra first. If you do not see it in your model picker yet, that is likely why.

Setting permissions to auto matters with a swarm. Many agents write many files at once, and stopping for approval on each one would slow the whole run down.

04Example 1: A Client System for a Web Design Agency

The first build came from a question inside the AI Profit Boardroom. A member who sells websites asked how to build their own client system instead of renting one. Julian called this build the client comb.

The prompt asked swarm mode to build the client comb with four workers, one module each:

  • Leads
  • Follow-ups
  • Bookings
  • Reviews

A shared shell joins the modules together. The prompt also said the verifiers must open the app in a real browser and reject anything that breaks.

The coordinator briefed nine agents. The shell worker laid down the design system first. Then the four module workers started writing code at the same time.

The leads module reached 22 kilobytes, while follow-ups, bookings and reviews all grew in parallel.

Then the inspectors checked each piece. The finished app showed follow-ups due today with the message already drafted, a booking calendar and a review request funnel. That is a working client system built from one prompt.

05Example 2: An AI Audit Kit for Consultants

The second build was for another member who is running their first AI audits this month, on site with two client businesses. Julian built an audit comb: an AI audit kit for an automation consultant.

A report worker pulled all the pieces into a client-ready audit. It includes a 30-60-90 day roadmap and a print-to-PDF button, so the consultant can hand it straight to the client.

The numbers from this run:

  • 477 kilobytes of code written.
  • About 32 minutes from start to finish.
  • Every piece passed the inspectors before it was sealed.

For the example agency used in the demo, the audit found 48 and a half hours a week that could be automated. That is the screen a consultant would walk their client through during an on-site visit.

This example shows why swarm mode suits bigger jobs. A full audit tool with several parts and a report would take one agent a long time step by step. With workers building side by side and a report worker joining it all at the end, the job finishes in one run.

06Example 3: A Weekly Update Desk for Coaches

The third build was for coaches and community owners. One of the busiest threads in the Boardroom that week was simply, how does anyone keep up? So Julian built a keeping up comb: a weekly update desk for a coach who runs an online community.

He seeded it only with the real Antigravity updates from that week, with nothing made up. Again, it used four workers:

  • Inbox collects the updates.
  • Triage sorts each update into use now, try later, or ignore for a coaching business.
  • Actions turns the use now pile into the top three moves for the week.
  • Member post writes a short community post about them, with a phone preview.

This is a useful pattern for anyone who runs a community. Instead of reading every update yourself, the desk sorts them, picks the moves that matter and drafts the post for your members. You stay in charge of what gets shared, but the reading and sorting is done for you.

07How the Verifier Agents Caught and Fixed Mistakes

The keeping up comb is the run where the inspectors really proved their value. Nobody had to tell them to look. They found problems, sent the work back and checked it again on their own.

Here is what happened, step by step:

  1. Verifier one passed 54 checks.
  2. It then sent the shell worker back to fix a missing file, and that piece lit up red on the board.
  3. Verifier two found two visual issues in the member post screen and sent that worker back.
  4. Both workers fixed their pieces.
  5. Both verifiers checked again.
  6. The final result was 57 out of 57 checks passed, with zero errors, and the build was sealed.

This is the main reason to use verifier agents. Without them, you would find the missing file or the layout bug yourself, after the run. With them, the swarm catches its own mistakes before you ever open the app.

The key is that the verifiers had clear rules in the prompt. They knew to test in a real browser and to reject anything that broke.

08Tips and Limits for Antigravity Swarm Mode

Julian shared four lessons from running these builds. They will save you time on your first swarm.

  • Be specific about who owns what. Give one module to each worker. That is what lets them build at the same time without stepping on each other.
  • Write the verifier rules into the prompt. For example: test in a real browser, no console errors, no overlaps, and data survives a reload. That is what made the checks real.
  • Expect heavy quota use. Every worker is its own agent, so a swarm uses far more than a single chat.
  • Use the Graph tab. After a run, scrub back through it to see who sent what to whom, and when.

The simple rule for picking a job: if it splits into three or four parts that do not depend on each other, it is a swarm job. If every part depends on the one before it, a single agent may be the better choice.

Because swarm mode is still in preview, expect it to change. Start with a small, clearly split project so you can learn how the coordinator and verifiers behave.

09Who Antigravity Swarm Mode Is For and What to Build Next

Antigravity swarm mode is for anyone who wants to build a full working tool from one prompt, without coding each part by hand. The three builds in this episode were for a web design agency, an automation consultant and a coach who runs a community.

Julian also suggested other swarm jobs for different kinds of businesses:

  • If you run a SaaS: an onboarding flow with welcome screens, a setup checklist and help docs, one worker each.
  • If you sell a course: a student portal with lessons, progress tracking and a question box.
  • If you run a store: a product page, a size guide and an FAQ, built side by side.

Each of these splits cleanly into separate parts, which is exactly what swarm mode is good at. The coordinator plans, the workers build in parallel, and the verifiers make sure it all works before you see it.

Julian shares all three swarm prompts, the hive and a step-by-step guide inside the AI Profit Boardroom, along with his latest Antigravity setups.