Google's Antigravity agent teams formalize the Weil conjectures and prove a 2013 open problem

Google’s agent teams just handed in a remarkable report card. On Friday, Google Research published results from Antigravity’s /teamwork, a system that coordinates teams of specialist Gemini 4 agents against long-open problems: a 2013 conjecture in information theory, landmark theorems in pure mathematics, molecular structures in chemistry, and a video decoder rewritten in a safer language.

Taken at face value, it is the strongest evidence yet that multi-agent systems can do sustained scientific work. The caveat: the most eye-catching numbers come from Google’s own account, and the code behind the biggest one hasn’t been published.

A conjecture from 2013, settled and checked by machine#

The headline result is the Courtade–Kumar conjecture, posed in 2013. It asks: when a random string of bits passes through a noisy channel, can you squeeze it down to a single bit while keeping more information than the laziest possible method — just copying one input bit? The conjecture says no.

Researchers from Google, CUHK, USC and Carnegie Mellon proved it in full generality with /teamwork’s Long Proof pattern, which splits a long argument across separate agents. The proof itself appeared as a preprint in September; what the November post adds is the machine-checked formalization — the whole argument translated into Lean 4, a language that lets a computer kernel verify every logical step. Google puts that formalization at 50 million lines of code.

That number deserves a closer look. An independent check of the public repository found the 50 million lines are real — but only about 239,000 are proof a person could read. The rest — some 44,000 files — is generated certificate data the Lean kernel checks mechanically. The line count measures how hard the checker was worked, not how much mathematics was done. The audit trail is nevertheless strong: the repository records a clean rebuild with a byte-identical final artifact.

The Weil conjectures, formalized without a human in the loop#

The more startling claim comes next. Google says a framework called /teamwork-LEAP, built with Google Cloud AI and paired with Gemini 4 Argon, “fully autonomously produced an end-to-end, machine-checked formalization of the Weil conjectures” — the landmark results on counting solutions of polynomial equations over finite fields, proved by Pierre Deligne in 1974. It spans about 1.8 million lines of Lean, produced across 28,800 subagent invocations.

Formalizing a theorem means turning its proof into code a proof checker verifies line by line; once mathematicians confirm the formal statement matches the original problem, the proof only compiles if every step holds. The same system, Google says, also produced the first formalization of a recent proof of Hilbert’s tenth problem over general rings of integers — more than 919,000 lines of Lean — plus first formalizations of the Birkhoff conjecture (open since 1927) and the 3D Bochner–Riesz conjecture (open 50+ years).

Here the verification story is thinner: Google’s post links to Wikipedia pages, with no repository, no paper and no statement file. A week earlier, Axiom published its own Weil formalization under an open license — about 213,000 lines, produced in less than a day by its own account. Until Google publishes its code, nobody outside the company can say what theorem it states, which axioms it uses, or whether it rebuilds.

The same machinery, pointed at chemistry and code#

The math is only part of the report. In chemistry, Google organized 30 specialized agents around a physics-informed spectral simulator to determine molecular structures from nuclear magnetic resonance data — where a single misplaced bond can separate a working drug from an inactive look-alike. It identified the exact structure 95% of the time on complex benchmark molecules, with the right answer in its top three every time — nearly double a single agent.

In software, the same approach rewrote libgav1, Google’s AV1 video decoder, from C++ into around 170,000 lines of Rust, with exact bit-for-bit parity. The first port ran about four times slower than the original. Given profiling tools and a mandate to maximize speed, the agents closed the gap: the Rust version now runs 20–30% faster than the original C++ and hand-optimized assembly.

The design idea is the same throughout: pair parallel attempts with dedicated “falsifiers” — the Lean proof checker, unmodified test suites, simulators, hardware counters — that decide which candidates actually work. Each agent gets only the files and tools its step needs, and critics audit results from a clean context, without seeing the original producer’s reasoning.

What to watch#

Google is making /teamwork and /boost available in Antigravity now, and has opened applications giving 10,000 academic researchers six months of free access to its highest usage tier. More proof patterns and specialized workflows are promised in the coming weeks.

Three things to watch: whether Google publishes the Weil formalization — the difference between a claim and a checkable result; whether the field’s scoreboard moves from line counts to verified artifacts; and whether the review side spreads: ahead of STOC 2027, authors of 578 submissions opted in to Deep Review feedback, and 90% of surveyed authors rated it mostly or almost all factual, with over 30% saying it caught substantive proof bugs.

The through-line is not that agents can now do mathematics. It is that a system built around verification — checkers that cannot be charmed, critics that never saw the first draft — can keep a thousand parallel attempts honest. The next year will show whether that holds outside Google’s own lab.

Sources: Google Research — Advancing science & engineering with Antigravity’s /teamwork (Nov. 13, 2026); AI Brief — Google says Gemini 4 agent teams solved open math problems and formalized the Weil conjectures (Nov. 14, 2026); Interesting Engineering — Google’s Antigravity agents formalise major mathematical proofs (Nov. 14, 2026); sysnote — Google’s /teamwork math claims, checked (Nov. 14, 2026). Prior coverage: Gemini 4 Carbon is now available to everyone — and Google’s /swarm turns one prompt into a team of agents.