Google's multi-agent Antigravity system formalises landmark maths proofs, researchers report
Google Research reported that teams of specialised AI agents in its Antigravity system formalised the Weil conjectures and other hard mathematical results, identified molecular structures and rewrote a video decoder, as the /teamwork tools open to outside researchers.
Story So Far
Teams of specialised AI agents working inside Google's Antigravity system have formalised landmark mathematical proofs, identified molecular structures and rewritten a video decoder, Google Research reported on November 13, as it opened the system's /teamwork tools to outside researchers and developers.
The headline result is the formalisation of the Weil conjectures, the landmark set of results on the number of solutions to polynomial equations over finite fields, proved by Pierre Deligne in 1974. Google said its Teamwork-LEAP system produced about 1.8 million lines of Lean code across 28,800 subagent invocations to formalise them. The system also reproduced a formalisation of a recent proof on finite-time blowup in the three-dimensional Navier-Stokes equations — a problem connected to one of the seven Millennium Prize Problems — and formalised the Courtade-Kumar conjecture, an information-theory problem first proposed in 2013, in Lean 4.
Beyond mathematics, Google applied the approach to molecular chemistry, where 30 specialised agents working with a physics-informed simulator identified molecular structures from nuclear magnetic resonance data with 95 percent top-one accuracy and 100 percent top-three accuracy on complex benchmark molecules. In systems engineering, the agents rewrote the company's libgav1 AV1 video decoder from C++ into about 170,000 lines of Rust, achieving exact output parity across a broad testing suite.
How we know?
Every figure here is Google's own, published in a research post by Google Fellow and vice-president Vahab Mirrokni and relayed by the trade press. No independent verification of the figures appeared in the reporting, and the Deep Review survey Google cites — 70 authors, 90 percent of whom rated the feedback mostly or almost all factual — is small and self-selected. The Weil conjectures formalisation is machine-checked in Lean, which makes it verifiable in principle; the results as published are Google's account of its own systems.
Original reporting
Independent corroboration
If the results hold up, the shift matters: multi-agent systems are moving from demonstrations to long-horizon research work, and the bottleneck moves from generating ideas to verifying them. The work was done with human experts throughout, except for the Teamwork-LEAP formalisation, which Google describes as fully autonomous — a distinction that will be tested as outside researchers get their hands on the tools.
External researchers' use of /teamwork is the next test. Independent attempts to reproduce the formalisations, and results on problems Google did not choose itself, will show whether this is a general advance or a showcase.
Keywords
Tags
- US
- Technology
- AI
- Research
