Google has expanded its Antigravity multi-agent system, utilizing its Gemini model-driven expert team to solve long-standing mathematical problems, verify complex proofs, and optimize software systems. The company stated that its Teamwork framework assists researchers in formalizing significant mathematical results, identifying molecular structures, and developing faster software implementations. This work combines human expertise with the collaboration of multiple agents, which independently explore solutions, test methods, and use computational tools to verify results.
One of the most notable achievements is the formalization of the Weil conjecture, a set of groundbreaking mathematical results concerning the number of solutions to polynomial equations over finite fields. Google’s Teamwork-LEAP system generated approximately 1.8 million lines of Lean code in 28,800 agent calls to formalize these results, which were initially proven by mathematician Pierre Deligne in 1974. Google also reported that the system reproduced the formalization of the latest proof regarding finite-time blow-up of the three-dimensional Navier–Stokes equations, a problem associated with one of the seven Millennium Prize Problems.
The question of whether smooth solutions to these equations can develop singularities in finite time has remained unresolved for about 90 years.
Google's Antigravity System Achieves Major Advances in Mathematics and Chemistry
The system also helped researchers resolve the Courtade–Kumar conjecture, an information theory problem first proposed in 2013. Researchers from Google and academic institutions developed a comprehensive proof using the framework, with the Gemini model assisting in deriving analytical bounds and generating certificates for computational verification. The proof was formalized in Lean 4, a programming language and proof assistant that allows computers to check mathematical arguments. Google’s approach relies on multiple agents working in parallel, assigning different tasks to each system to challenge and verify proposed solutions, aiming to reduce errors that may arise when a single model generates long chains of reasoning.
Each expert agent receives only the documents and tools necessary to perform its task. Independent reviewers and auditors examine the results without relying on the original agents' reasoning. In addition to mathematics, Google has applied this method to molecular chemistry. A group of 30 expert agents collaborated with physical information simulators to identify molecular structures from nuclear magnetic resonance (NMR) data, achieving a 95% accuracy rate for first place and 100% accuracy for the top three, targeting complex benchmark molecules.
The framework has also achieved results in systems engineering. Google stated that its agents rewrote the company's libgav1 AV1 video decoder from C++ into approximately 170,000 lines of Rust code, achieving complete output consistency across a broad test suite. The initial rewrite ran about four times faster than the original implementation. When agents were provided with performance analysis tools and tasked with improving performance, the Rust version became 20% to 30% faster than the original C++ implementation on x86_64 and ARM processors.
In another project, a two-week initiative rewrote the remaining C++ code in the Zircon microkernel into approximately 800,000 lines of memory-safe Rust, passing kernel unit tests and being able to boot on ARM64 hardware.
Google also reported improvements to its internal infrastructure. Its intelligent scheduling workflow reduced preemption of TPU workloads by 24% on certain hardware types while maintaining service level objectives and cluster utilization. The company is offering Teamwork and its iterative coding workflow /boost through Antigravity, with 10,000 academic researchers invited to apply for six months of free high-usage access. The company stated that additional expert workflows, including formal proof tools, will be made available in the future.