Google opens Antigravity's /teamwork agent teams to outside developers, with a checker for every claim
Quick answer: In a Nov 13, 2026 Google Research post, Vahab Mirrokni says outside researchers and developers can use /teamwork and /boost in Antigravity with Gemini 4 models today. /teamwork runs many agents in parallel and pairs them with “dedicated falsifiers” such as the Lean 4 proof checker or a project’s unmodified tests. A formal-proof pattern and a code-rewrite pattern are still listed as upcoming. (Google Research, Nov 13)

How it works
- Parallel tries, independent checkers. Many candidate solutions run at once, and each is tested by something that can break it: Lean 4, unmodified test suites, simulators, web search or hardware counters. (Google Research)
- Split context. Each agent gets only the files and tools for its step. Critics and auditors review from a clean context, without the producer’s reasoning. (Google Research)
- /boost is the same idea packaged as a single sequential solve-and-verify loop. (Google Research)
What is available when
- Today: /teamwork and /boost with Gemini 4 models in Antigravity, plus Deep Review. Google says it has opened applications for 10,000 academic researchers to get six months of free access to its highest usage tier. (Google Research)
- Upcoming (“soon” or “in the coming weeks,” no date given): the formal proof pattern, a code-rewrite pattern, and more tutorials on the Antigravity blog. (Google Research)
The results Google reports (all Google’s own, self-reported, from the Nov 13, 2026 post)
- Math: /teamwork-LEAP produced a machine-checked formalization of the Weil conjectures of about 1.8 million lines of Lean, across 28,800 subagent invocations. Google says it did this fully autonomously.
- Chemistry: 95% top-1 exact match on complex benchmark molecules in Google’s own NMR structure-identification test, using 30 agents and a spectral simulator. Google says that nearly doubles single-agent performance.
- Code rewrite: about 800,000 lines of safe Rust added in a two-week campaign on the Zircon microkernel, passing all kernel unit tests and booting on arm64 hardware.
- Google’s own infrastructure: two shipped scheduling changes cut TPU workload preemption by 24% on some hardware types, a measure inside Google’s Borg scheduler. No baseline is given.
What is not verified
We have not seen independent review of any of these results. A Lean check confirms a proof matches the statement as written, not that the statement matches the intended theorem. Lines of code measure size, not quality, and passing tests is only as strong as the tests. (Our reading; a similar caution appears in coverage by Times of AI.)
FAQ
Is it free? Not generally. Access follows Antigravity plans; the free six-month offer is for 10,000 academic researchers who apply.
Which models? Gemini 4 Argon, Carbon and Hydrogen are named in the post.
Does this mean AI solved open math problems on its own? Google says the work was done “in collaboration with human experts” and describes researchers framing the questions.
Sources
- Google Research, “Advancing science & engineering with Antigravity’s /teamwork” (Nov 13, 2026): https://research.google/blog/advancing-science-engineering-with-antigravitys-teamwork-feature/
- Google Antigravity Blog, “Teamwork: when AI becomes a research partner”: https://antigravity.google/blog/teamwork-when-ai-becomes-a-research-partner/
- Antigravity docs, Teamwork: https://antigravity.google/docs/teamwork/
- Times of AI, caveats on the results (Nov 2026): https://timesof.ai/2026/11/google-antigravity-teamwork-falsifiers-weil-lean-zircon-rust
— AltTab Dad
