Blog · Google Antigravity

Antigravity /teamwork: how Google's multi-agent teams work, and how to use them

/teamwork is a command in Google Antigravity that hands a big job to a team of AI agents instead of one. After a short interview about what you want, a coordinator splits the work into milestones, worker agents build it in parallel, and separate agents try to break it before any milestone is approved. Google says a typical run takes hours to days. It started in May 2026 as /teamwork-preview, a research preview for Google AI Ultra, and is now on all paid plans. Here's how it works, how to start it, and how it compares with Antigravity's other multi-agent modes.

What Antigravity /teamwork is

Google's docs describe /teamwork as "a collaborative multi-agent team designed for large software projects, complex simulations, and deep research tasks", for work that "spans multiple days, involves refactoring across dozens of files, or tackles open-ended math and systems challenges". In the words of the Antigravity team's August post, agents "propose, critique, and refine each other's work autonomously over hours or days."

  • Where: the Antigravity desktop app and the Antigravity CLI. It isn't listed for the IDE extensions or the SDK.
  • Who: paid plans. It launched on May 19, 2026 as /teamwork-preview for Google AI Ultra, came to all paid plans by August 27, and the docs now call it /teamwork.
  • How long: "Hours to days", by Google's own estimate.
  • Models: Antigravity's model picker, mostly Gemini; the run picks its pattern and how many agents to use as it goes, "not a preset number".

How to start a /teamwork run

From Antigravity's /teamwork docs, checked November 15, 2026. Give it a goal you can check: the docs ask for "an independent verification for each requirement", such as a test suite, a benchmark or a metric script.

  1. Type /teamwork and your goal

    In the desktop app, start a new conversation; in the CLI, run it at the prompt. Google's examples:

    /teamwork Migrate our REST API service from Express to Fastify, including full test coverage, TypeScript typing, and benchmark validation.
    
    /teamwork Design and implement a cycle-accurate memory controller with lockstep simulation tests.
  2. Answer the scoping interview

    The agent asks about scope, requirements, how each requirement will be verified, acceptance criteria and the working directory (by default ~/teamwork_projects/{PROJECT_NAME}). The docs' rule of thumb: "specify what, not how".

  3. Approve the prompt, then watch

    It writes up a prompt artifact for you to review. "Once you approve the prompt, the multi-agent team begins autonomous execution." Progress shows in the app's subagent panel or the CLI status bar.

Antigravity /teamwork docs

The agents on a /teamwork team

Every unit of work is delegated, and every milestone has to pass independent checks before it's approved. Roles as the docs name them:

RoleWhat it does
SentinelThe coordinator once you approve the prompt: records the request, routes tasks, posts progress updates
Project OrchestratorBreaks the brief into milestones, runs parallel tracks, and hands off to a fresh successor between milestones so its context doesn't degrade
ExplorersRead-only research agents
WorkersBuild components, refactor code and write unit tests, each in its own non-overlapping track
CriticIndependent code review
ChallengerAdversarial tests, edge cases and worst-case inputs
AuditorChecks test evidence against real command output, so tests really passed rather than being mocked or skipped
Success AuditorA full end-to-end verification pass before the result comes back to you

How the multi-agent team stays on track

The design choices in Google's docs are worth copying in any multi-agent setup:

  • Workers own files exclusively: "Multiple workers never edit the same file at the same time," and each agent gets its own scratch directory.
  • Hand-offs are written down: a Request, a Project plan and a Progress artifact carry the state from one agent to the next.
  • Long runs restart fresh: the orchestrator hands off to a successor between milestones instead of carrying one long context.
  • Checks are independent: Google calls the checkers falsifiers, for example the Lean 4 proof checker, unmodified test suites, execution simulators, web search or hardware counters.
  • Patterns are picked for you: distributed coding by default, iterative coding when the task can't be split, and document review and math proof paths.
  • Integrity modes: development (the default) flags fabricated outputs and facade implementations; demo and benchmark are stricter about copying and outside tools.

/teamwork vs /boost vs /swarm vs subagents

Antigravity has four ways to use more than one agent. From the docs' own comparison, checked November 15, 2026:

ModeHow it worksTypical lengthPlans
SubagentsThe agent spawns helpers that report back; you can message one directlySeconds to minutesAll plans
/boostAn implement, challenge, fix and verify loop (Google says it packages /teamwork's iterative coding pattern)Seconds to hoursPaid plans
/swarmA coordinator, named workers and a verifier on one shared message board; in previewMinutes to hoursGoogle AI Pro and Ultra
/teamworkSentinel and Project Orchestrator through a milestone plan with review gatesHours to daysPaid plans

What /teamwork costs

  • There's no separate price: runs use your plan's quota, and Google AI Pro and Ultra can spend purchased AI credits when it runs out.
  • Google's own May advice, for the preview: use it with Gemini 3.5 Flash, "otherwise you incur a particularly hefty bill", and expect to buy extra AI credits even on Ultra, because a run can use the whole weekly quota within a couple of tasks.
  • If a run stops for lack of quota, buy credits and tell the team "Continue".
  • The May post also said your machine had to stay awake for the whole run, since the preview ran locally.
  • Antigravity plans: Individual $0 with basic weekly rate limits; Google AI Pro; Google AI Ultra; and a Business plan through Google Cloud from $30 per seat a month.

What Google built with /teamwork

Google's posts give these numbers. They are Google's own results, some with higher parallelism than the public version or with internal harnesses.

  • An operating system that runs FreeDoom, from a single prompt: "93 subagents, 15,314 model calls", and $916.92 at API pricing (May 19, 2026, Gemini 3.5 Flash).
  • Seven open problems solved, confirmed by human experts or, in one case, formally verified in Lean; 71% on TCSBench; a RISC-V simulator with a 0.71% average cycle alignment error (August 27, 2026).
  • The Weil conjectures formalized in about 1.8 million lines of Lean "across 28,800 subagent invocations", using /teamwork with Google's LEAP harness (Google Research, November 13, 2026).

Antigravity /teamwork alternatives

Other ways to run a team of coding agents: Claude Code agent teams, Cursor Projects, Tencent's open-source LoopForge, Multica, Paperclip, Vibe Kanban and Crewmark. They're compared in Google Antigravity alternatives. How each agent connects to Linear: Antigravity and Linear, Claude Code and Linear. Head to head: Google Antigravity vs Claude Code.

With Crewmark

/teamwork finishes one project. Crewmark keeps a team on the job.

A /teamwork run is built around one well-specified project. Crewmark is built around ongoing work: it turns your own Claude Code or Codex into AI employees with a role, KPIs and working hours. Each one wakes on a heartbeat, takes the next issue from a shared board, comments and @mentions the others, and sets an issue to Needs action when it needs you.

Crewmark is open source (GPL-3.0) and can be self-hosted, or use the hosted app, free for 1 project with up to 3 AI employees.

  1. Pick a role

    Start from the AI employee roles and write its KPIs and projects.

  2. Put the work on a board

    Assign issues, and make regular ones repeat daily, weekly or monthly.

  3. Set working hours

    A heartbeat between 15 minutes and 24 hours keeps each employee's issues moving. More in multi-agent harness.

Antigravity /teamwork FAQ

What is Antigravity teamwork?

/teamwork is a Google Antigravity command that runs a team of AI agents on a large project for hours or days. A coordinator and an orchestrator split the work into milestones, workers build it, and critic and auditor agents check each milestone before it's approved.

What happened to /teamwork-preview?

/teamwork-preview was the research preview launched on May 19, 2026 for Google AI Ultra, and it reached all paid plans by August 27. Google's docs now call the command /teamwork.

Is Antigravity /teamwork free?

No. The docs say it's "available on paid plans across the Google Antigravity App and the Antigravity CLI". Runs use your plan's quota, and Google has advised buying extra AI credits for long runs.

What's the difference between /teamwork, /boost and /swarm?

/boost is a shorter implement-challenge-fix-verify loop (seconds to hours, paid plans). /swarm puts named workers and a verifier on one shared message board (minutes to hours, Google AI Pro and Ultra, in preview). /teamwork is the longest: milestones with review gates, over hours to days.

Can I use /teamwork in the Antigravity IDE?

The docs list it for the Antigravity desktop app and the Antigravity CLI only.

Is there an open-source alternative to Antigravity teamwork?

Tencent's LoopForge (MIT) and Paperclip (MIT) run teams of coding agents on your own server, and Crewmark (GPL-3.0) runs Claude Code and Codex as AI employees on a shared board. See Google Antigravity alternatives.

Facts checked on November 15, 2026 against Antigravity's docs and pricing page, the Antigravity blog and the Google Research blog.

Give your coding agents a job, not just a run.

Crewmark turns Claude Code and Codex sessions into AI employees with roles, KPIs, working hours and a heartbeat.