Verdict: On November 13, 2026, Google Research announced that Antigravity's /teamwork multi-agent system proved the 2013 Courtade-Kumar conjecture in full generality with an end-to-end Lean 4-verified proof, while /teamwork-LEAP formalized the Weil conjectures autonomously in about 1.8 million lines of Lean across 28,800 subagent invocations. The same post opens the machinery to everyone: /teamwork, /boost and Deep Review are live in Antigravity, with six months of free top-tier access for 10,000 researchers. AI research output now counts only when a proof kernel compiles - and Google made that harness a product.
TL;DR (last verified November 15, 2026)
- The 2013 Courtade-Kumar conjecture is proved in full generality: no Boolean function beats a dictator, verified end-to-end in Lean 4.
- The Weil conjectures got an autonomous, machine-checked formalization: about 1.8 million lines of Lean, 28,800 subagent runs.
- First formalizations shipped for Hilbert's tenth problem over general rings of integers, a 3D Navier-Stokes blowup proof, the Birkhoff conjecture and the 3D Bochner-Riesz conjecture - with one gap found and fixed.
- Production wins too: a Rust AV1 decoder 20-30% faster than the C++ original, an 800,000-line microkernel rewrite, and TPU kernels cutting memory use up to 185x.
- The tools are open today, with a free-researcher program attached.
What the Courtade-Kumar proof settles
The conjecture is about squeezing information through a noisy wire. Take a random n-bit string, flip each bit independently with probability p, then let a function g summarize the corrupted input to a single bit. A dictator function, which copies one input bit, carries 1 - H2(p) bits, where H2 is binary entropy. Courtade and Kumar conjectured in 2013 (original conjecture) that no Boolean function does better in any dimension.
Thirteen years later that is a theorem. Researchers from Google, CUHK, USC and CMU used /teamwork's Long Proof pattern, which implements the Stellar Colosseum harness, to prove it in full generality (arXiv 2609.24931, listed September 24, 2026). The entire proof, including all numerical certificates, is formally verified in Lean 4 end-to-end; Google's post puts the formalization at 50 million lines of code (Google Research) - an artifact no human reviewer could certify by hand, which is the point.
The Weil conjectures, with no human in the derivation loop
The LEAP formal-verification framework, paired with Gemini 4 Argon, produced an end-to-end, machine-checked formalization of the Weil conjectures - the results on counting solutions of polynomial equations over finite fields that culminated in Pierre Deligne's 1974 proof (Wikipedia) - spanning about 1.8 million lines of Lean across 28,800 subagent invocations, fully autonomously (Google Research).
A cluster of firsts landed alongside, all from the same primary source:
- Over 919,000 lines of Lean for the first formalization of a recent proof of Hilbert's tenth problem over general rings of integers.
- An independent reproduction of an end-to-end formalization of a recent proof of finite-time blowup for the 3D Navier-Stokes equations with smooth forcing, part of one of the seven Millennium Prize Problems, open for roughly 90 years (Wikipedia).
- The first formalization of the Birkhoff conjecture, a dynamical-systems problem open since 1927.
- The first formalization of the 3D Bochner-Riesz conjecture, open in harmonic analysis for more than 50 years, where the system found and fixed a gap in a recently released informal proof.
One nuance: formalizing a modern proof certifies the argument; it does not settle the open problem. The Navier-Stokes regularity question remains open, and what LEAP reproduced is a machine-checkable version of a recent partial proof.
How /teamwork actually works
Two failure modes motivated the design. A single model generation is unpredictable, so /teamwork pairs parallel exploration with dedicated falsifiers - the Lean 4 proof checker, unmodified test suites, execution simulators, web search, hardware counters - that decide which candidates survive (Google Research). And long-horizon agents lose track of details as attention disperses, so context is partitioned by construction: each specialist gets only the files and tools its step needs, coordinators hand off to fresh successors, and critics and auditors verify from clean contexts without the producer's reasoning.
Workflows are declarative, self-improving natural-language patterns; a router sends each prompt to the pattern that fits, and the iterative coding pattern ships standalone as /boost. We compared orchestration-only products in our Cursor Projects Multiplayer vs Antigravity Swarm and Claude Science workbench reviews; the step here is that orchestration now ships with verification as a first-class citizen.
Production results beyond mathematics
The same falsifier-driven loop produced engineering results with real consequences:
- Chemistry: 30 specialized agents with a physics-informed simulator identified molecular structures from NMR data at 95% top-one and 100% top-three accuracy on benchmark molecules (Google Research), extending the agentic chemistry line in Antigravity Cheminformatics vs El Agente Farmaco.
- libgav1, Google's AV1 decoder, was rewritten from C++ into about 170,000 lines of Rust with 100% bit-exact parity across a broad conformance suite. The port initially ran about 4x slower; after the agents got profiling tools and a speed mandate, it finished 20-30% faster than the original C++ and hand-optimized assembly on x86_64 and ARM (Interesting Engineering).
- A two-week campaign rewrote the remaining C++ in Google's Zircon microkernel into roughly 800,000 lines of memory-safe Rust, passing kernel tests and booting on ARM64 (Interesting Engineering).
- Custom TPU kernels for graph neural networks, built with /teamwork on Argon, cut high-bandwidth-memory pressure by up to 185x and sped up forward and backward passes by 8x to 15x (Antigravity blog).
- Algorithmic theory moved too: an Argon harness proved and Lean-verified near-optimal sample bounds for adaptive data analysis, closing a gap open since 2014 and yielding the first nearly space-optimal streaming algorithm against adaptive adversaries (project proof PDF, arXiv 2611.08699); a new construction settled a 2014 quantum question with the first n-bit Boolean functions needing strictly more than n/2 quantum queries at constant error (arXiv 2611.09809); and Stellar Colosseum delivered the first single-pass cut-sparsification algorithm with near-optimal memory, update and reconstruction time, resolving a FOCS 2022 open question (arXiv 2611.03604).
- In a survey of Deep Review users, over 30% said it caught substantive proof bugs or mathematically significant typos, and 14% found technical gaps that took more than an hour to fix (Google Research).
What opens today, and where this goes
Everything described is live: /teamwork and /boost run inside Antigravity on Gemini 4 Argon, Carbon and Hydrogen (we benchmarked the price/performance trade in Gemini 4 Hydrogen vs GPT-6.1 Terra), Deep Review includes two free weekly calls (Google Research), and a form hands 10,000 academic researchers six months of free top-tier access. The formal proof pattern is listed as an upcoming addition.
Three things follow for the practical reader. If you draft research, the falsifier pipeline is the acceptance layer: write the Lean certificate or test suite before you trust the draft. If you run engineering teams, multi-day agentic campaigns now deliver measurable wins - a faster codec, a memory-safe kernel, cheaper TPU kernels - and the model-versus-platform question we tracked in Antigravity vs Cursor vs Claude Code shifts toward whoever owns verification tooling. And if you are evaluating AI for serious work, demand machine-checkable artifacts: the delta between the model said so and the kernel compiled is now the difference between a claim and a result.
That is the structural shift: AI research outputs have acquired a referee, the referee is a compiler, and Google handed the pipeline to the public.
FAQ
Q: What is Google Antigravity's /teamwork?
A: Multi-agent orchestration in Antigravity: declarative workflow patterns pair parallel exploration with falsifiers, and context is partitioned so each agent sees only what its step needs.
Q: Did AI solve the Navier-Stokes problem?
A: No. The system reproduced a Lean formalization of a recent finite-time-blowup proof under smooth forcing; the Millennium Prize question about smooth solutions remains open.
Q: Why does Lean 4 verification matter so much here?
A: Lean turns a proof into code a kernel checks line by line; the proof only compiles if every logical step holds, so machine-checked output replaces trust in the model's wording.
Q: Can you use /teamwork today, and what does it cost?
A: Yes - /teamwork and /boost are live in Antigravity on Gemini 4 models, Deep Review includes two free weekly calls, and Google is accepting applications from 10,000 researchers for six months of free top-tier access.
Q: What exactly was proved about the Courtade-Kumar conjecture?
A: That for a random bit string passed through a noisy channel, no Boolean function preserves more information in a single output bit than the dictator function that copies one input bit, with the bound equal to 1 minus the binary entropy; the full proof is Lean-verified end-to-end.
Last verified: November 15, 2026. Written by Sham Uddin, founder of aitecharchive.com (information science, machine learning systems engineering), with AI assistance; every figure above was checked against the primary sources linked inline. Corrections and methodology: /pages/how-we-work.
Top comments (0)