← Slices

Google's Antigravity /teamwork Formalizes Weil Conjectures in Lean

Google's Antigravity /teamwork Formalizes Weil Conjectures in Lean

Key takeaways

  • Google Research says its Antigravity /teamwork multi-agent system produced an end-to-end, machine-checked Lean formalization of the Weil conjectures, about 1.8 million lines long.
  • That formalization took 28,800 subagent invocations, and the post does not put a price on any of the runs.
  • Lean checks logic, not intent: a human mathematician still has to confirm the formal statement matches the original problem.
  • For solo operators, the transferable idea is the verifier: every agent output should face an independent check it cannot argue its way past.
  • These are Google's own results, reported in Google's own blog post, so treat them as claims until outside experts have examined them.

Google Research published a post on November 13, 2026, under the byline of Vahab Mirrokni, Google Fellow and VP at Google Research, describing what Antigravity's /teamwork multi-agent system has produced on open problems. The headline result is an autonomous, end-to-end formalization of the Weil conjectures, the landmark results on counting solutions to polynomial equations over finite fields that culminated in Pierre Deligne's 1974 proof. Google says it spans about 1.8 million lines of Lean, produced across 28,800 subagent invocations.

Put plainly: a team of software agents wrote a very long formal proof of one of the landmark results of modern algebraic geometry, and the referee is a machine. That is either the start of a serious research assistant or an extremely expensive way to discover where your proof checker runs out of patience. Probably both.

What /teamwork actually does

The post describes two problems with current models that /teamwork is built to handle. First, a single attempt is unpredictable, so the system runs parallel exploration and pairs it with falsifiers: tools that check which candidate answers actually work, such as the Lean 4 proof checker, unmodified test suites, simulators, web search or hardware counters. Second, long tasks lose the thread as context grows, so /teamwork partitions context by design. Each agent receives only the files and tools it needs for one step, coordinators hand work to fresh successors, and critics and auditors verify results from clean contexts without seeing the producer's reasoning.

Review gates over writers: the same logic behind /teamwork's critics and auditors. We made this argument earlier in why every agent needs an editor.

The Weil run was paired with Gemini 4 Argon, according to the post. Google's collaborators also used other Gemini 4 models, including Gemini 4 Hydrogen, which we covered in our Gemini 4 Hydrogen analysis.

The other results Google is putting on the table

Beyond the Weil conjectures, the post reports several results:

  • A 919,000-line Lean formalization of a recent proof of Hilbert's tenth problem over general rings of integers, which Google describes as the first formalization of that proof.
  • A full-generality proof of the 2013 Courtade–Kumar information-theory conjecture, produced with researchers from CUHK, USC and CMU. Google says the entire proof was formalized in Lean 4 at about 50 million lines of code.
  • In chemistry, 30 specialized agents paired with a physics-informed spectral simulator reached 95% top-1 exact match and 100% top-3 accuracy on complex benchmark molecules from NMR data. Google says this nearly doubled single-agent performance.
  • In systems engineering, a /teamwork rewrite of libgav1, Google's AV1 video decoder, from C++ into roughly 170,000 lines of Rust. The port began about four times slower than the original. After a profiling-driven push, Google says it ran 20 to 30 percent faster than the C++ and hand-optimized assembly on x86_64 and ARM, with 100% bit-exact parity on its conformance suite.
  • Internally, Google says a Borg agentic scheduling workflow shipped two changes that cut TPU workload preemption by 24% on some hardware types.

Google's framing deserves a caveat. Every one of these is a Google result, reported by Google. The post names academic collaborators but does not describe any external peer review, and the Courtade–Kumar figure in particular is the kind of number you want to see in the repository before you repeat it.

The part that matters: the checker

Lean is a proof assistant. It does not care whether a theorem is interesting; it cares whether each step follows. The post is candid about the catch: once human mathematicians confirm that the formal theorem statement matches the original problem, the proof compiles only if every logical step holds. The machine does the proving, and humans still do the reading. The hard part moves from "is the proof right?" to "did we write down the right question?"

Google also says some /teamwork capabilities are already available to external developers and researchers through Antigravity and Deep Review, while others, including the formal proof pattern, are coming in the weeks ahead. It has also opened applications for free six-month access to its highest usage tier for 10,000 academic researchers.

What it means for solo operators

Nobody reading this is going to Lean-check a marketing funnel. The architecture underneath the result travels fine, though. Three ideas from the post translate to a one-person shop:

  • Give every agent a falsifier. Tests, reconciliations, a rendered-page check, a crawl. If nothing can say no, the agent will say yes, cheerfully and at scale.
  • Scope the context. An agent doing one step with the files it needs beats one carrying your entire business in its head. We covered this in what breaks in long-running agents after step three.
  • Keep a human at the statement. Google's own post says the human's job is to confirm the question. Yours is the brief, the acceptance criteria and the definition of done. For when to stop an agent mid-task, see the human-in-the-loop trigger.

Our take

I have spent more than 25 years in digital marketing, and I have watched enough "revolutionary" launches to hold my applause until the receipts arrive. So here is my honest read. Formal verification is the most underrated idea in agentic AI right now, and /teamwork's best move is not the 1.8 million lines. It is the decision to make the checker the referee and the model the player who has to satisfy it. That is the same reason we keep saying an agent without an editor is a very fast intern with no supervisor.

My skepticism points in two directions. First, fifty million lines for a single information-theory proof is the sort of figure that makes me want the repository before I believe the adjectives. Second, the bottleneck Google admits to, human confirmation of the formal statement, does not disappear. That is the expensive part, and it is where people who can read a theorem keep earning their fees. Mathematicians, you may stand down from panic; the robot still needs you to write the question down.

As a marketer, I will say the uncomfortable thing out loud: most AI output in our industry would fail a checker that never gets tired. If your content, reports or ad copy have no falsifier, you do not have an AI workflow. You have a faster way to be wrong. For the business case behind the same shift in math, read our look at OpenAI's push into verifiable mathematics.

Sources

Google Research: Advancing science & engineering with Antigravity's /teamwork (November 13, 2026)

Frequently asked questions

What did Google announce about Antigravity /teamwork?

Google Research said /teamwork multi-agent workflows produced results across mathematics, chemistry, formal verification and systems engineering, and that some of these capabilities are being made available to outside researchers and developers.

Did a machine prove the Weil conjectures?

Google says its /teamwork-LEAP system produced a machine-checked Lean formalization of the Weil conjectures, about 1.8 million lines long. Human mathematicians still have to confirm that the formal statement matches the original problem.

Can developers use /teamwork today?

According to Google, Gemini 4 models with /teamwork and /boost are available to external users in Antigravity and Deep Review now. Some features, including the formal proof pattern, are due in the coming weeks.

What should a solo operator take from this?

Build an independent check into every agent workflow, such as tests, reconciliations or audits the agent cannot talk its way past. Scope each agent's context to the step at hand, and keep a human responsible for the brief.

Related reading

Book a 30-min discovery →