Google Says Its Antigravity /teamwork Agents Machine-Checked the Weil Conjectures and Cracked Long-Open Problems
A Google Research post credits Gemini 4 multi-agent workflows with a full Lean proof of a 2013 information-theory conjecture, a 1.8-million-line formalization of the Weil conjectures and an 800,000-line Rust kernel port. Google is opening the tools to outside researchers, with free access for 10,000 academics.
Published
Last updated / Fact-checked on
Illustrative photo of calculus written on a blackboard. It does not show any of the proofs or code described in Google's post.
Google Research has published its most sweeping claim yet for what teams of AI agents can do on hard technical work. In a post dated Nov. 13, Google Fellow Vahab Mirrokni described results from /teamwork, a multi-agent feature inside Google's Antigravity coding environment, across mathematics, chemistry, systems software and machine-learning engineering. The headline items are a full machine-checked proof of an information-theory conjecture open since 2013, an autonomous formalization of the Weil conjectures in the Lean proof language, and a two-week campaign that ported the rest of a Google microkernel from C++ to memory-safe Rust.
Google stresses that the work was done "in collaboration with human experts." It is also making the tools available to outside users, which is what turns a research showcase into a product story.
At a glance
| Item | Google's claim |
|---|---|
| Courtade–Kumar conjecture (2013) | Proved in full generality; proof formalized in Lean 4 at 50 million lines |
| Weil conjectures | End-to-end Lean formalization of about 1.8 million lines over 28,800 subagent calls |
| Zircon microkernel | About 800,000 lines of safe Rust added in a two-week campaign; passes kernel unit tests and boots on arm64 |
| libgav1 AV1 decoder | Ported to about 170,000 lines of Rust, bit-exact, and 20–30% faster than the C++ original after tuning |
| Access | /teamwork and /boost in Antigravity and Deep Review today; free top-tier access for 10,000 academics for six months |
How /teamwork is built
The post argues that two weaknesses of today's language models drive the design. First, a single attempt at a hard problem is unreliable, so /teamwork runs several attempts in parallel and pairs them with what Google calls falsifiers: tools that can mechanically reject a wrong answer, such as the Lean 4 proof checker, unmodified software test suites, execution simulators, web search or hardware counters.
Second, models lose track of details over multi-day work as their context windows fill up. /teamwork splits context deliberately. Each specialized agent sees only the files and tools it needs for its step, coordinators hand work to fresh successors, and critic and auditor agents check results from a clean context, without seeing the reasoning of the agent that produced them. Workflows are written as natural-language patterns, and the system routes each request to the pattern that fits, from distributed coding to long-proof and document review. A sequential solve-and-verify loop is packaged separately as /boost.
The mathematics
The flagship result is the Courtade–Kumar conjecture, a 2013 problem in information theory. In Google's description, it says that when a random bit string passes through a noisy channel, no single-bit summary of the input preserves more information than simply copying one input bit. Researchers from Google, CUHK, USC and CMU worked with /teamwork's long-proof pattern. Google says Gemini first found counterexamples to a more general version of the conjecture, the CUHK team then reduced the problem to a 14-variable optimization with a key lower bound, and Gemini 4 Argon and earlier Gemini models produced most of the remaining bounds and interval-arithmetic certificates, extending the proof from balanced to unbalanced functions without human input. The team formalized the entire proof in Lean 4.
Google lists a long run of other results: near-optimal bounds in adaptive data analysis that it says close a gap open since 2014, a deterministic combinatorial algorithm answering the main open question from a STOC 2024 paper on Boolean matrix multiplication, a quantum query-complexity construction settling a 2014 question, and progress on Tarjan's 1985 Deque Conjecture for splay trees, verified in Lean.
On formal verification, Google says a framework called LEAP, built with Google Cloud AI and paired with /teamwork and Argon, "fully autonomously" produced a machine-checked formalization of the Weil conjectures, the results on counting solutions of polynomial equations over finite fields that Pierre Deligne completed in 1974. It also reports the first formalization of a recent proof of Hilbert's tenth problem over general rings of integers, at more than 919,000 lines, and a formalization of a recent proof of finite-time blowup for the 3D Navier–Stokes equations with smooth forcing. Google says the system found and fixed a gap in a recently released informal proof of the 3D Bochner–Riesz conjecture while formalizing it.
The engineering
The software results may matter more to most readers. Google says a /teamwork rewrite pattern, which schedules each module only after its dependencies compile and tests every piece against the original, unmodified test suite, completed the port of the Zircon microkernel to Rust after an earlier 25% port. The same approach rewrote libgav1, Google's AV1 video decoder. The Rust version started out about four times slower than the C++ original; after the agents were given profiling tools and told to maximize speed, Google says it ended up 20–30% faster than the C++ and hand-optimized assembly on both x86-64 and ARM.
Other claims include custom TPU kernels for graph neural networks, a 9x9 Go engine trained from scratch that gained roughly 2,600 Elo in under eight days, and production wins inside Google, including a 24% cut in TPU workload preemption on some hardware from changes to the Borg scheduler.
What is actually available
External scientists and developers can use Gemini 4 models with /teamwork and /boost in Antigravity, and Google's Deep Review manuscript checker, today, according to the post. Google has opened applications for 10,000 academic researchers to get six months of free access to its highest usage tier. The formal-proof pattern and the Rust rewrite pattern are described as coming later, with tutorials promised on the Antigravity blog in the coming weeks. Deep Review users get two free weekly calls.
Google also reports that authors of 578 submissions to the STOC 2027 theory conference opted in to Deep Review feedback, and that in an ongoing survey of 70 authors, 90% rated the feedback mostly or almost all factual and over 30% said it caught substantive proof bugs or significant typos.
Why it matters
The pattern in Google's post is that the impressive results come where a machine can say no. Lean will not compile a false proof; an unmodified test suite will not pass a broken port; a hardware timer does not care how confident a model sounds. That makes /teamwork less a claim that models now reason flawlessly than a claim that, wrapped in enough checking, they can grind through problems too long for any one attempt.
The caveats are real. Every figure here is Google's own, published on its research blog, and most of the mathematical results will need scrutiny from the wider community, including whether each formal statement matches the problem it claims to settle. Google itself notes that human mathematicians must confirm that a formal theorem statement matches the original problem before a compiled proof means anything. Watch for the underlying papers and Lean repositories, for independent mathematicians' reviews of the Weil and Navier–Stokes formalizations, and for whether outside researchers using the public tools can reproduce results of this kind on their own problems.
How we reported this
DailyMech read Google Research's Nov. 13 post in full. All results, figures and availability details come from that post and are Google's own claims. DailyMech did not review the underlying papers, proofs or code. DailyMech contacted no one for this article.
Published November 14, 2026. Facts checked against the cited sources by DailyMech's AI writer.
Sources
- Advancing science & engineering with Antigravity's /teamwork, Vahab Mirrokni, Google Research blog, Nov. 13, 2026 (fully read).
DailyMech used AI tools to research and draft this article. Facts were checked against the cited sources by DailyMech's AI writer, not by a human editor.