Google Research has published a long account of what the multi-agent mode of its Antigravity coding platform, /teamwork, has produced over recent weeks — and the list reads less like a product update than like the results section of a research paper.

At the centre is the Courtade–Kumar conjecture, a 2013 problem in information theory asking whether any compression of a noisy channel's input down to a single bit can preserve more information than a 'dictator function' that simply copies one input bit. Researchers from Google, CUHK, USC and CMU used /teamwork's Long Proof pattern running on Gemini 4 Argon to prove the conjecture in full generality. Gemini surfaced counterexamples to a generalised version; a new candidate function reduced the problem to a 14-variable optimisation, and the agents developed most of the analytical bounds and interval-arithmetic certificates — including extending the proof from balanced to unbalanced functions without human input. The whole result was then formalised in Lean 4, producing 50 million lines of machine-checkable code.

The same harness, paired with a formalisation framework called LEAP, produced machine-checked formalisations of the Weil conjectures — about 1.8 million lines of Lean across 28,800 subagent invocations — plus first-time formalisations of the Birkhoff conjecture (open since 1927) and the 3D Bochner–Riesz conjecture (open for more than 50 years). It also reproduced an end-to-end formalisation of finite-time blowup for the 3D Navier–Stokes equations, part of a Millennium Prize problem.

Outside mathematics, Google says a /teamwork rewrite pattern — which schedules each module only after its dependencies compile, linking every leaf module against the original unmodified test suite — ported the remaining C++ of the Zircon microkernel to roughly 800,000 lines of memory-safe Rust that boots on arm64 hardware. The same pattern rewrote Google's libgav1 AV1 video decoder into about 170,000 lines of Rust with 100% bit-exact parity. After being given profiling tools and a mandate to maximise speed, Google says the Rust decoder ended up 20–30% faster than the original C++ and hand-optimised assembly on both x86_64 and ARM.

There are chemistry results too: 30 specialised agents in a self-correcting loop with a physics-informed spectral simulator reached 95% top-1 exact match on molecular structure elucidation from NMR data, roughly doubling single-agent performance.

Google says /teamwork and the related /boost coding loop are available in Antigravity today, and that applications have opened giving 10,000 academic researchers six months of free access at its highest usage tier. Big claims in AI research deserve the same scrutiny the blog applies to the proofs: these are Google's own accounts, several results rest on agent-generated proofs that human collaborators edited and verified, and nothing substitutes for independent peer review.