JOURNAL / 2026.11.15
Google announces a Weil formalization with /teamwork: what LEAP lets us check
Google reports a large-scale mathematical formalization and expands its agent workflows. LEAP's method explains how it preserves progress and reviews dependencies; assessing the evidence requires separating the announced result, the published system and what users can reproduce.
The Weil conjectures already had a proof. The news is turning deep mathematics into an object that a proof checker can inspect. Google Research announced on November 13 that /teamwork-LEAP, paired with Gemini 4 Argon, produced their complete formalization in Lean: approximately 1.8 million lines through 28,800 subagent invocations. This is a result reported by Google; we have not recompiled that formalization. The announcement still lists its formal proof pattern among upcoming Antigravity updates.
The scale deserves attention, although counting lines or calls does not measure elegance, cost or ease of review. My reading is that the potential advance lies in making a chain of dependencies manageable when a team would struggle to maintain it by hand. To assess that advance, I would ask to see the final statement, its assumptions, the libraries used and an independent compilation. A large number cannot replace that process.
The dates matter. Teamwork was already offered as a preview on paid plans in August, with agents proposing and criticizing strategies. Also, the Courtade–Kumar conjecture manuscript, highlighted again in the announcement, appeared in September. November's publication brings results together and expands the reported scope; it does not turn every cited work into a discovery from this week.
Preserve what is proved, while keeping unfinished work visible
The LEAP paper, revised on November 12, describes a memory of goals and dependencies. A goal can have alternative routes; each route requires all its lemmas to be resolved. The system preserves verified results and can reuse them. The compiler checks formal steps, while a model-based reviewer tries to reject unproductive decompositions. Its main experiments use Gemini 3.1 Pro: they are not a direct evaluation of the Weil case with Argon.
This makes it possible to ask a more useful question than how many agents participate: what counts as progress? Splitting a problem can multiply tasks without moving closer to a solution. I would want a record distinguishing proved steps from outstanding obligations and explaining what each result unlocks. That distinction would help decide where to spend more computation and where to change strategy.
There is a public route to exploring the method. The LEAP-CLI repository offers integration with coding agents and describes a final output without sorry, Lean's placeholder for an unfinished proof, checked by the kernel using standard axioms. This describes how the program works; it is not an independent audit of every announced result. Access to the code also does not supply the inference budget of a research campaign.
For a team wanting to try it, I would start with a lemma whose statement an expert can review. I would fix definitions and assumptions, set a budget and have the final file checked in a separate environment. I would measure human preparation and review work alongside runtime and consumption. My expectation would be to learn which parts of formalization become less manual, rather than extrapolate from one example how long any theorem would take.
The /teamwork documentation confirms access on paid plans and an initial phase for agreeing on requirements and checks. It describes critics, agents looking for failures and auditors. Those functions organize review, but a role's name does not determine the strength of the evidence. Agreement between models, a test suite and a formal proof answer different questions.
In my view, the most important opportunity is to build results that outlive the agent that produced them. A researcher should be able to inspect and reuse a dependency without inheriting the entire conversation behind it. That possibility makes this publication worth following. The decisive test will be how much independent work its artifacts enable, and how much effort remains necessary to confirm that the formal statement expresses the mathematical question we intended to solve.
Primary sources consulted on November 15, 2026. We have not run /teamwork or LEAP or reproduced their results. The evaluation proposals and assessment are our own analysis.