Google Teamwork Multi-Agent System Formalizes Weil Conjectures
Google has expanded its Antigravity multi-agent system into Teamwork, a Gemini-powered framework that pairs parallel specialist agents with independent critic and reviewer agents. It formalized the Weil conjectures in roughly 1.8 million lines of Lean code across about 28,800 sub-agent calls, reproduced a formal proof on finite-time blowup for the 3D Navier-Stokes equations, and delivered verifiable results in molecular structure identification and a C++ to Rust rewrite of its libgav1 decoder.
Related Entities
Full Report
IT之家 November 14 news, Google announced further expansion of its Antigravity multi-agent system, using multiple specialized agents powered by Gemini models to form the Teamwork system, which is claimed to solve "long-standing unsolved mathematical problems," "verify complex mathematical proofs," and optimize the performance of various software systems.
For example, Google has successfully formalized a proof of the Weil conjectures. This is a set of important mathematical results concerning the number of solutions to polynomial equations over finite fields, originally proved by mathematician Pierre Deligne in 1974. Google's Teamwork-LEAP system completed the formalization of the relevant mathematical results through approximately 28,800 sub-agent calls, generating a cumulative total of about 1.8 million lines of Lean code.
Google also stated that the corresponding system recently successfully reproduced a formal proof concerning the finite-time blowup problem of the three-dimensional Navier–Stokes equations. It also helped researchers solve the Courtade–Kumar conjecture.
From a technical perspective, Google's Teamwork system mainly relies on multiple agents working in parallel, with an independent system arranged to question and verify the proposed solutions, in order to reduce errors that a single AI model may produce during long-form reasoning. Each specialized agent can only access the files and tools required to complete the task, and then independent questioners and reviewers check the results without relying on the original agent's reasoning process.
In addition to mathematical research, Google has also applied this technology to the field of molecular chemistry. A team of 30 specialized agents, combined with a simulator incorporating physical knowledge, used nuclear magnetic resonance (NMR) data to identify molecular structures. In complex molecule benchmarks, the system achieved a Top-1 accuracy of 95% and a Top-3 accuracy of 100%.
In the field of systems engineering, Google also demonstrated application results of multi-agent collaboration. The company stated that its agent team rewrote its own libgav1 AV1 video decoder from C++ to Rust, ultimately generating about 170,000 lines of code and achieving output results completely identical to the original version in a wide-ranging set of tests.
Currently, Google is making the Teamwork system and an iterative programming workflow called /boost available to users through Antigravity, and plans to provide 10,000 academic researchers with free access to the highest usage quota for 6 months.