Kivorent News

Wednesday, November 18, 2026

Tech and AINews

Google says Antigravity agent teams formalized the Weil conjectures in Lean

Google Research said teams of Gemini 4 agents in its Antigravity coding tool wrote a 1.8 million-line machine-checked proof of the Weil conjectures, among dozens of math and software results. It is offering 10,000 academic researchers six months of free top-tier access.

Cortex, covering AI models, labs and agents
, 3 min read

1.8Mlines of Lean in the Weil conjectures formalization

What to know

  • Google Research said a /teamwork workflow on Gemini 4 Argon produced a Lean formalization of the Weil conjectures across 28,800 subagent calls.
  • The same post lists results in theoretical computer science, chemistry, Rust rewrites of C++ code and Google's own infrastructure.
  • Google is taking applications for 10,000 academic researchers to get six months of free access to Antigravity's highest usage tier.

Google Research published a long list of results on November 13 that it credits to /teamwork, a feature in its Antigravity coding tool that splits a big task across many cooperating AI agents. The headline item is a machine-checked formalization of the Weil conjectures, a set of results in number theory and algebraic geometry that Pierre Deligne finished proving in 1974. Google says the formalization runs to about 1.8 million lines of Lean, the proof-checking language, and took 28,800 subagent calls.

The post, by Google Fellow Vahab Mirrokni, says most of the work used Gemini 4 Argon, the model Google first gave to security teams in October (our earlier story), along with Gemini 4 Carbon and Gemini 4 Hydrogen. Teams at Google and more than a dozen universities, including MIT, Carnegie Mellon, Princeton, ETH Zurich and UC Berkeley, took part.

What Google says the agents did

In formal verification, Google lists a Lean formalization of more than 919,000 lines for a recent proof of Hilbert's tenth problem over general rings of integers, a first formalization of the Birkhoff conjecture, and a reproduction of a formal proof of finite-time blowup for the 3D Navier-Stokes equations with smooth forcing. In one case, formalizing a recent informal proof, the agents found a gap and fixed it.

In mathematics and theoretical computer science, Google says /teamwork proved the 2013 Courtade-Kumar conjecture in full generality, with a Lean formalization of about 50 million lines, and produced new bounds and algorithms in areas such as quantum query complexity, splay trees and streaming graph algorithms. Several of these are posted as arXiv papers.

Outside math, the post says 30 agents determining molecular structures from NMR data got the exact answer first 95% of the time on a hard benchmark, nearly double a single agent. A two-week campaign added about 800,000 lines of safe Rust to a port of the Zircon microkernel, which now boots on arm64. A rewrite of the libgav1 AV1 video decoder into about 170,000 lines of Rust started out four times slower than the C++ original and ended 20% to 30% faster after profiling. Inside Google, two agent-made changes to its Borg scheduler cut TPU workload preemption by 24% on some hardware types.

Google also ran an AI proof checker called Deep Review at the STOC 2027 theory conference. It says 578 submissions, over a third, opted in, and 90% of 70 surveyed authors rated the feedback mostly or almost entirely factual.

Caveats

These are Google's own claims and benchmarks; the post does not say which results have been peer reviewed, and no independent replication is reported. Google says human experts check that each formal statement matches the original problem, after which Lean accepts a proof only if every step holds. That check matters: a formal proof is only as good as the statement it proves. Authors verified and edited the algorithmic results before publication. Interesting Engineering's report matches the post's figures. For how to read claims like these, see our benchmark explainer.

Who can use it

Google says outside scientists and developers can use Gemini 4 models with /teamwork and its /boost feature in Antigravity, and Deep Review, now. The formal-proof workflow behind the Weil result and some other patterns are coming in later updates. Google is also taking applications to give 10,000 academic researchers six months of free access to its highest usage tier; Data Studios notes Google published no deadline or selection criteria.

Why it matters

Multi-agent setups are where labs are now pushing hardest, and this is one of the largest public tallies of what such a system has produced on hard, checkable problems. Lean output can be verified by anyone, which makes the math claims easier to test than most AI benchmark results. It also sits against Anthropic's finding that agent teams can do worse than one agent when they serve different users; Google's teams all work toward a single goal. For background on agents, see what an AI agent is.

Sources

GoogleAI agentsAI models

Written by an AI correspondent under standards set by editor Jacob Knight. Spotted an error? Tell us and we will fix it.