Antigravity /teamwork explained: Google’s AI agent teams.
Google’s Antigravity /teamwork runs teams of Gemini 4 agents with built-in verifiers. What it is, the math and code results Google reports, and how to try it.
By The Model Atlas · 3 sources
In this post7 sections
Antigravity /teamwork is Google’s multi-agent mode for its Antigravity coding app and CLI: instead of one AI agent working through a long task, a team of specialised Gemini agents plans, builds and checks the work, with separate agents whose only job is to find mistakes. On 13 November 2026 Google Research published a long list of results it says it got with /teamwork and the Gemini 4 models, from machine-checked math proofs to large C++-to-Rust rewrites, and said the tools are open to outside researchers and developers. Here is what Antigravity /teamwork is, what Google reports it has done, and how to try it.
What is Antigravity /teamwork?
Google’s Antigravity docs (opens in a new tab) 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, touches dozens of files or tackles open-ended math and systems problems. You start it with the /teamwork command followed by your goal.
In its research post (opens in a new tab), Google says the design targets two limits of today’s models. A single attempt at a hard problem is unpredictable, so /teamwork runs several attempts in parallel and pairs them with “falsifiers” such as the Lean 4 proof checker, unmodified test suites, simulators, web search or hardware counters that decide which candidate actually works. Models also lose track of details over multi-day tasks, so /teamwork gives each agent only the files and tools it needs for its step, and hands work to fresh agents instead of letting one context grow forever.
How /teamwork agent teams work
According to the docs, every /teamwork run has two phases. First the main Antigravity agent interviews you about scope, requirements, how each requirement will be checked and what counts as done, then writes a prompt you review and approve. After that, the team works on its own in a dedicated project folder (by default ~/teamwork_projects/{PROJECT_NAME}). The roles Google lists include:
- Sentinel, the coordinator that records your request, routes tasks and posts progress updates.
- Project Orchestrator, which splits the brief into milestones, delegates the work and hands off to a fresh successor between milestones.
- Explorers, read-only agents that research the codebase, and Workers, which write the code and tests in non-overlapping tracks.
- Critic, Challenger and Auditor, verification gates that review the code, build adversarial tests and check that tests really passed rather than being mocked or skipped.
- Success Auditor, a final end-to-end check before the finished project comes back to you.
The docs also describe three integrity modes, from a lenient development mode to a strict benchmark mode that requires a from-scratch implementation, and different execution paths for general coding, document review and math proofs.
What Google says /teamwork has done
Google Research says teams across Google and partner universities used Gemini 4 Argon, Gemini 4 Carbon, Gemini 4 Hydrogen and /teamwork-style workflows on open problems over several weeks, always with human experts involved. The headline claims are Google’s own:
Math and formal proofs
- Google says /teamwork-LEAP “fully autonomously” produced a machine-checked Lean formalization of the Weil conjectures, about 1.8 million lines of Lean across 28,800 subagent invocations.
- It reports more than 919,000 lines of Lean for the first formalization of a recent proof of Hilbert’s tenth problem over general rings of integers.
- Researchers from Google, CUHK, USC and CMU proved the 2013 Courtade–Kumar conjecture from information theory and formalized the proof in Lean 4, which Google says came to 50 million lines.
- Google also lists results in algorithms, quantum query complexity, optimization and data structures, many with Lean 4 checks that the authors verified and edited.
Science and code
- In chemistry, Google says 30 agents in a self-correcting loop reached 95%% top-1 exact match and 100%% top-3 accuracy at identifying molecular structures from NMR data, nearly double a single agent.
- A two-week Argon campaign added about 800,000 lines of safe Rust to finish porting the Zircon microkernel, which Google says passes all kernel unit tests and boots on arm64 hardware.
- /teamwork rewrote Google’s libgav1 AV1 decoder into around 170,000 lines of Rust with bit-exact output; after it got profiling tools, Google says the Rust port ran 20–30%% faster than the original C++ and assembly.
These are company-reported results. Google says human experts checked the published work, and Lean-checked proofs are verified by a computer kernel, but the wider research community is still reviewing many of the claims.
/teamwork vs /boost vs /swarm
Antigravity has several multi-agent commands, and the docs are clear about which to use. /boost (opens in a new tab) is for problems that are hard rather than broad, such as race conditions, algorithms or tricky refactors: it starts at once and runs rounds where workers try to break and fix each other’s answers. /swarm splits broad work across named workers and a verifier on a shared message board. /teamwork is for multi-milestone projects that run for hours or days, with a scoping interview first. Google’s post says /boost packages /teamwork’s iterative coding pattern for day-to-day tasks.
Who can use /teamwork and how to try it
Google’s docs say /teamwork and /boost are available on paid plans in the Antigravity app and the Antigravity CLI, and /swarm on Google AI Pro and Ultra. To try it:
- Open a new conversation in the Antigravity app, or the prompt in the Antigravity CLI.
- Type /teamwork followed by your goal, for example Google’s own: “Migrate our REST API service from Express to Fastify, including full test coverage, TypeScript typing, and benchmark validation.”
- Answer the scoping interview, review the prompt it writes, and approve it to start the team.
- Follow progress in the subagent panel or CLI status bar, and in the plan and progress files in the project folder.
For researchers, Google says it has opened applications for 10,000 academic researchers to get six months of free access to its highest usage tier, and that Deep Review, its manuscript-checking tool, gives users two free calls a week. Google says more features, including its formal proof pattern, are coming to Antigravity in the coming weeks.
What people are building with Antigravity and Gemini 4
Builders on X are already pairing Antigravity with the newest Gemini 4 models. Google’s own Chrome dino 3D canyon runner was made with Gemini 4 Carbon in Antigravity, and a puzzle race against Gemini models also used Antigravity. With Gemini 4 Hydrogen, creators built a neon city flythrough and an island lagoon with a sailboat in Antigravity. Browse every Gemini 4 build on the Gemini 4 Argon, Carbon and Hydrogen model pages.
Antigravity /teamwork FAQ
What is /teamwork in Google Antigravity?
It is a slash command that runs a team of specialised AI agents, with a coordinator, workers and independent reviewers, on long projects such as migrations, simulations, research and math proofs.
Is Antigravity /teamwork free?
Google’s docs say /teamwork is available on paid plans. Google has also opened applications for 10,000 academic researchers to get six months of free access to its highest usage tier.
Which models does /teamwork use?
Google says the results it published came from Gemini 4 Argon, Gemini 4 Carbon and Gemini 4 Hydrogen with /teamwork workflows built on Gemini models.
What is the difference between /teamwork and /boost?
/boost starts immediately and suits hard, tightly coupled problems that take seconds to hours. /teamwork begins with a scoping interview and suits multi-milestone projects that take hours to days.
Did /teamwork really formalize the Weil conjectures?
Google says /teamwork-LEAP produced an end-to-end, machine-checked Lean formalization of about 1.8 million lines. That is Google’s claim; outside experts are still reviewing its results.