# Google opens Antigravity /teamwork multi-agent workflows to outside researchers, 10,000 free seats

Canonical: https://theagenticweekly.com/news/google-antigravity-teamwork-open-problems · Evidence: Reported · First seen: 2026-11-13 22:16 UTC · Source published: 2026-11-13 20:47 UTC · Brief published: 2026-11-14 06:06 UTC · Updated: 2026-11-14 06:06 UTC · Author: Agentic Weekly

> Google Research said on Nov 13, 2026 that "External scientists and developers can use Gemini 4 models with /teamwork and /boost in Antigravity and Deep Review today," and opened applications for 10,000 academic researchers to get six months of free access to its highest usage tier.

## What changed

- Antigravity's /teamwork multi-agent workflows and the /boost command are available to external users with Gemini 4 models, alongside Deep Review. Google says a formal proof pattern and other specialized workflows will come to Antigravity in later updates.
- Google opened applications that give 10,000 academic researchers six months of free access to its highest usage tier.
- /teamwork partitions context by design: each specialized agent gets only the files and tools for its step, coordinators hand work to fresh successors, and critics and auditors re-check results from clean contexts.
- Candidate answers are checked by dedicated falsifiers such as the Lean 4 proof checker, unmodified test suites, execution simulators, web search or hardware counters, rather than by the model grading itself.
- /boost packages /teamwork's iterative solve-and-verify coding loop for day-to-day tasks. A /teamwork rewrite pattern for large code migrations, which schedules each module only after its dependencies compile, is described as soon to be available.

## The results Google attributes to /teamwork

Selected figures from the Google Research post (Google's claims, not independently checked)

| Area | Claimed result |
| --- | --- |
| Formal math | Weil conjectures formalized in about 1.8 million lines of Lean across 28,800 subagent invocations |
| Chemistry | 30 agents with a spectral simulator: 95% top-1 exact match, 100% top-3 on benchmark molecules |
| Systems code | About 800,000 lines of safe Rust from a two-week campaign completing a Zircon microkernel port |
| TPU kernels | 4.09x geometric-mean speedup over reference implementations using DuoDiscover |
| Paper review | 578 STOC 2027 submissions opted in to Deep Review; 90% of 70 surveyed authors rated feedback mostly or almost all factual |
| Google production | 24% reduction in TPU workload preemption on some hardware types via Borg agentic scheduling |

## What agent builders can take from the design

Our reading: the transferable idea is the pairing of parallel attempts with an outside checker, plus strict context partitioning. Google's two stated reasons are that a single model attempt is unpredictable and that attention drifts over multi-day tasks as context grows. Its answer is not a bigger context window but narrower ones: each agent sees only what its step needs, and reviewers start fresh without the producer's reasoning. Teams building their own long-running agents can copy that pattern with any model, as long as the task has a real falsifier such as a test suite, a proof checker or a benchmark.

The limit is the same one Google names: the gains come where results can be verified mechanically. Open-ended tasks without a checker, such as product writing or strategy, are not what these numbers describe.

## In the source's words

> "External scientists and developers can use Gemini 4 models with /teamwork and /boost in Antigravity and Deep Review today."
> (Google Research blog, https://research.google/blog/advancing-science-engineering-with-antigravitys-teamwork-feature/)

> "This formalization spans about 1.8 million lines of Lean, produced across 28,800 subagent invocations."
> (Google Research blog, on the Weil conjectures formalization, https://research.google/blog/advancing-science-engineering-with-antigravitys-teamwork-feature/)

## What we have not verified

- We have not run /teamwork or /boost, so the speedups and accuracy figures are Google's own claims, not results we reproduced.
- The mathematical results are reported by Google and its academic co-authors; we did not check the Lean formalizations or papers.
- The post does not state Antigravity pricing or which plans include /teamwork beyond the academic free-access program.
- Several features, including the formal proof pattern and the rewrite pattern, are described as coming later, without dates.

## Timeline

- Source published: 2026-11-13 20:47 UTC
- First seen by our watcher: 2026-11-13 22:16 UTC (via Google Research)
- Brief published: 2026-11-14 06:06 UTC

## Sources

- [Google Research: Advancing science & engineering with Antigravity's /teamwork](https://research.google/blog/advancing-science-engineering-with-antigravitys-teamwork-feature/) (primary source) (fetched 2026-11-14)

## Related

- [All news briefs](https://theagenticweekly.com/news)
- [New AI models this week](https://theagenticweekly.com/models/this-week)

## How to cite

Agentic Weekly, "Google opens Antigravity /teamwork multi-agent workflows to outside researchers, 10,000 free seats", published 2026-11-14 06:06 UTC, https://theagenticweekly.com/news/google-antigravity-teamwork-open-problems