Google is advancing its innovative Antigravity system with a new development called Teamwork, which leverages a group of specialized AI agents powered by the Gemini model to tackle intricate problems, including verifying mathematical proofs and optimizing software.
Intelligent Assistants for Science and Technology
Teamwork has demonstrated remarkable capabilities in scientific research. The system recently succeeded in formalizing the proof of the Weil conjectures, fundamental results concerning the number of solutions to polynomial equations over finite fields. This proof, first presented by Pierre Deligne in 1974, was formalized by the Teamwork-LEAP system using approximately 2.88 thousand subsystem calls, generating around 1.8 million lines of code in the Lean programming language.
Furthermore, Teamwork replicated the formalization of the proof for the existence of explosive solutions to the 3D Navier-Stokes equations in finite time. The system also assisted researchers in resolving the Courtade–Kumar conjecture.
How Teamwork Operates
The core innovation of Teamwork lies in the parallel operation of multiple AI agents. The system also incorporates independent mechanisms for checking and verifying proposed solutions. This approach minimizes the risk of errors that can arise when a single AI model tackles lengthy and complex reasoning. Each specialized agent is granted access only to the files and tools necessary for its assigned task. The results of their work are then reviewed by independent “critics” and “reviewers” who do not rely on the original agents’ reasoning process.
Applications in Chemistry and Software Development
The Teamwork technology extends beyond mathematics. In molecular chemistry, a team of 30 specialized agents, utilizing simulators with integrated physical knowledge, successfully identified molecular structures based on Nuclear Magnetic Resonance (NMR) data. In challenging tests on molecular databases, the system achieved a Top-1 accuracy of 95% and a Top-3 accuracy of 100%.
In systems engineering, Google has also showcased the effectiveness of multi-agent collaboration. A team of agents rewrote the company’s libgav1 AV1 video decoder from C++ to Rust. The outcome was approximately 170,000 lines of code that produced identical results compared to the original version across a wide range of tests.
Accessibility and Future Prospects
Google is now opening access to the Teamwork system and its iterative programming/boost workflow for users via the Antigravity platform. The company plans to offer free access with certain usage limits for six months to 10,000 academic researchers, underscoring Google’s commitment to collaboration and accelerating scientific progress through advanced AI technologies.
