What Google says it did
Google Research published a post on 13 November describing results from /teamwork, a multi-agent feature in its Antigravity development environment. The headline claim concerns the Weil conjectures, a set of statements about counting solutions to polynomial equations over finite fields that occupied number theorists for decades and were closed by Pierre Deligne in 1974.
A system the company calls /teamwork-LEAP produced what Google describes as an end-to-end, machine-checked formalisation of those conjectures, culminating in Deligne’s proof. “This formalization spans about 1.8 million lines of Lean, produced across 28,800 subagent invocations,” the post says. Google says the system produced it autonomously.
What a Lean formalisation actually certifies
Lean is a proof assistant: a program that will only accept a proof if every step follows from the axioms and lemmas available to it. A formalisation that compiles is therefore a strong guarantee — but a guarantee about the theorem as formalised, not about whether the statement written down is the one mathematicians mean by the Weil conjectures. Checking that correspondence is human work, and Google’s post does not say it has been done by anyone outside the company.

That matters because the compiler is, for now, the only referee on offer. Google does not say which of these results have been peer reviewed or submitted for publication, and the line counts, the subagent counts and the claims of autonomy are all the company’s own.
The other results in the post
The Weil work is not the only claim. Google says the same system synthesised more than 919,000 lines of Lean to formalise a recent proof of Hilbert’s tenth problem over general rings of integers, which it describes as the first such formalisation. It says the system independently reproduced an end-to-end formalisation of a recent proof of finite-time blowup for the three-dimensional Navier-Stokes equations with smooth forcing — a problem the post notes is one of the seven Millennium Prize Problems.
Separately, Google says researchers used a “Long Proof” workflow pattern to prove the Courtade-Kumar conjecture, open since 2013, in full generality, and then formalised that proof in Lean 4, producing 50 million lines of code. The post also lists work on the Birkhoff conjecture, open since 1927, the three-dimensional Bochner-Riesz conjecture, and regularity of infinity-harmonic functions, alongside results in combinatorics, data structures and algorithms.
How the system is built

The architecture Google describes is less exotic than the numbers suggest. /teamwork runs parallel exploration and then filters candidates through verifiers — the Lean 4 proof checker in the mathematics work, unmodified test suites in the engineering work. It partitions context so each agent sees only what it needs, and routes each prompt to a workflow pattern suited to the task. The verifier, not the model, is what decides whether a candidate survives.
Google also reports a survey of its Deep Review manuscript-review tool with 70 respondents, in which 90% rated the feedback as mostly or almost all factual and 90% said they would use it again. That is a small sample, self-selected, and reported by the vendor.
Who gets to use it
External scientists and developers can use Gemini 4 models with /teamwork and /boost in Antigravity now, and Deep Review users get two free calls a week. Applications are open for a programme giving 10,000 academic researchers six months of free access to the highest usage tier.
That programme is the thing to watch. If the Lean artefacts are published and mathematicians outside Google start reading the statements these agents formalised, the 1.8 million-line figure becomes checkable rather than merely large.