November 15, 2026AgentsResearchFramework

Google's /teamwork Proves Weil in Lean. The Grader Did the Real Work

1.8 million lines of Lean, 28,800 subagent calls, and the Weil conjectures formally proved "fully autonomously," by Google's own count. That is the headline of a Google Research post dated Nov 13 (no timestamp on the page; it reached Hacker News at 13:20 UTC on Nov 14) on what the Antigravity /teamwork harness has done with Gemini 4 Argon, Carbon and Hydrogen: https://research.google/blog/advancing-science-engineering-with-antigravitys-teamwork-feature/

The list keeps going. The Courtade-Kumar conjecture, open since 2013, is proved in full generality with a 50-million-line Lean proof, after human collaborators at CUHK, USC and CMU reduced it to a 14-variable problem. More than 919,000 lines of Lean cover Hilbert's tenth problem over rings of integers. The first formalization of 3D Bochner-Riesz found and fixed a gap in a recently released informal proof. On the engineering side, a two-week Argon campaign picked up an earlier 25% port of the Zircon microkernel and rewrote the remaining C++, adding about 800,000 lines of safe Rust, passing every kernel unit test and booting on arm64. The libgav1 AV1 decoder was ported to roughly 170k lines of Rust, bit-exact, first about 4x slower than the C++, then 20 to 30% faster than C++ and hand-tuned assembly once the agents got profiler access. And a Deep Review tool read 578 STOC 2027 submissions; in a survey of 70 authors, over 30% said it caught substantive proof bugs.

The model is not the interesting part. Every result has a grader the agents can't talk their way past: the Lean kernel, unmodified test suites, bit-exact conformance, simulators, hardware counters. Google calls them dedicated falsifiers. Context is partitioned by construction, coordinators hand off to fresh successors, and critics audit from a clean context without the producer's reasoning, so the auditor can't inherit the producer's blind spots. Put simply, generation is cheap and the checker decides what counts. Lexifina's Lean model of the US tax code (posted Nov 10, 138 drafting faults found) runs the same recipe at a much smaller scale.

The product half shipped the same day. Antigravity v2.32.0, dated Nov 13, adds Dynamic Workflows: type /run-workflow and the agent writes a short Python script with parallel() and pipeline() that spawns other agents, each in its own conversation, each returning a JSON-schema result. You approve the script, it runs in the background, and you can save it and rerun it: https://antigravity.google/docs/dynamic-workflows. Two days earlier, Claude Code's Code Mode had Claude write JavaScript to make its tool calls. Two labs, one direction: orchestration is moving out of the chat and into code.

Caveats matter here. These are Google's numbers, human co-authors verified, simplified and edited the math per the post, and the formal-proof pattern is still listed as upcoming in the product. Still, the lesson travels. If your domain has a Lean kernel or a conformance suite, this harness design works today. If it doesn't, building the falsifier is the first job, not picking the model.
← Previous
Nerfed Daily: 2026-11-14
Next β†’
llama.cpp Hits 1.0: The Local Stack Finally Promises Not to Break
← Back to all articles

Comments

Loading...
>_