Subscribe
0%
Illustration for: Google's Antigravity agents: Courtade–Kumar proof, 1.8M-line Weil formalization, Zircon in Rust. What's verified

Google's Antigravity agents: Courtade–Kumar proof, 1.8M-line Weil formalization, Zircon in Rust. What's verified

Google Research says teams of Gemini agents in Antigravity's /teamwork helped prove a 2013 information-theory conjecture, machine-checked the Weil conjectures in Lean and rewrote large C++ codebases in Rust. Some results are formally verified. Others are preprints or single runs reported by Google.

Listen to this article0:00 / 5:13

Google Research published a long list of results on November 13 from /teamwork, the multi-agent mode in its Antigravity coding tool. The headline claims: a full proof of the Courtade–Kumar conjecture, open in information theory since 2013, and an autonomous Lean formalization of the Weil conjectures about 1.8 million lines long. The post, by Google Fellow Vahab Mirrokni, also covers Rust rewrites of Fuchsia's Zircon microkernel and the libgav1 AV1 decoder.

This matters for two reasons. First, the results show AI agents running for hours or days, with separate agents checking each other's work. Some of that output is machine-checked math, which is much harder to fake than a benchmark score. Second, Google is opening these tools up. It says /teamwork, /boost and Deep Review are available with Gemini 4 models in Antigravity now. It's also taking applications from 10,000 academic researchers for six months of free access to its top usage tier.

The tool itself isn't new. Google's Antigravity blog described /teamwork in August, when it was credited with seven smaller open-problem results. The November post is a much bigger batch.

The math results, and how they were checked

Courtade–Kumar (2013). This conjecture says that no single-bit summary of a random bit string sent over a noisy channel preserves more information than simply copying one of the input bits (a "dictator" function). Google says researchers from Google, CUHK, USC and CMU proved it in full generality. In Google's account, Gemini found counterexamples to a generalised version, and the CUHK team reduced the problem to a 14-variable optimisation. Gemini 4 Argon and earlier models then produced most of the analytical bounds and interval-arithmetic certificates.

The arXiv preprint lists seven authors: three from CUHK, plus Google researchers including Mirrokni and David Woodruff (also at CMU) and Adel Javanmard (also at USC). Its abstract says the whole proof, including the numerical certificates, was formally verified in Lean end to end. Google's post puts the Lean output at 50 million lines. That makes the proof checkable by machine, but it hasn't been peer reviewed yet.

Weil conjectures. Pierre Deligne finished proving these in 1974. Google says its LEAP formal-verification framework, paired with Gemini 4 Argon, produced a fully autonomous, end-to-end machine-checked formalization. It reports about 1.8 million lines of Lean across 28,800 subagent calls. No new mathematics is involved here. The feat is in translating a famously deep proof into code a computer can check.

Other formalizations. Google lists:

  • a formalization of a recent proof of Hilbert's tenth problem over general rings of integers (more than 919,000 lines)
  • what it calls the first formalizations of the Birkhoff and 3D Bochner–Riesz conjectures. For Bochner–Riesz, Google says the work found and fixed a gap in a recent informal proof.
  • a reproduction of a formalization of a recent finite-time blowup proof for 3D Navier–Stokes with smooth forcing

That last item doesn't solve the Millennium Prize problem, and Google doesn't claim it does.

New theory results. Google also lists:

  • near-optimal bounds in adaptive data analysis, Lean-verified
  • a quantum query complexity result settling a 2014 question
  • a deterministic Boolean matrix multiplication algorithm answering an open question from STOC 2024
  • several streaming and sketching bounds, mostly as arXiv preprints

Engineering claims

Project What Google reports
Zircon microkernel → Rust ~800,000 lines of safe Rust in a two-week campaign; passes all kernel unit tests, boots on arm64 hardware
libgav1 AV1 decoder → Rust ~170k lines with bit-exact output on a conformance suite; started 4x slower, ended 20–30% faster than the C++ and hand-tuned assembly after profiling
NMR structure solving 30 agents plus a spectral simulator: 95% top-1 and 100% top-3 accuracy on "complex benchmark molecules"
9x9 Go engine ~2,600 Elo gained in under eight days, to 1,885 Elo, between KataGo at 16 and 32 visits
TPU kernels 4.09x geometric-mean speedup across 17 JAXBench kernels
Borg scheduling 24% less TPU workload preemption on some hardware types

Google says the Rust rewrite pattern and the formal-proof pattern are "coming soon" for outside users.

What to keep in mind

  • Verification varies. The Courtade–Kumar proof, the Weil formalization and some theory results are Lean-checked. Many other items are preprints or linked code with no peer review yet.
  • Self-reported. Every figure in the table comes from Google. We found no outside replication. The NMR benchmark's size and makeup aren't described, and the Go claim rests on rating estimates plus a single showcase game.
  • Humans in the loop. Researchers set the goals, made key reductions (as CUHK did for Courtade–Kumar) and decided what to accept. Google frames the agents as research partners, not autonomous discoverers, except for the Weil formalization.

Who should care: mathematicians and theoretical computer scientists, who can now check these Lean files and apply for the free access tier. Also teams with big C or C++ codebases, if the Rust rewrite pattern ships and holds up outside Google's own code.

Sources

  1. Google Research — Advancing science & engineering with Antigravity's /teamwork research.google
  2. arXiv — A Proof of the Most Informative Boolean Function Conjecture arxiv.org
  3. Interesting Engineering — Google's Antigravity agents formalise major mathematical proofs interestingengineering.com
  4. Kantan News — Google Antigravity multi-agent system ushers in a new era in mathematics and software kantan.news
  5. Google Antigravity Blog — Teamwork: When AI becomes a research partner (August) antigravity.google

#Google#Antigravity#Gemini#AI agents#mathematics#Lean#Rust