Google Research releases Antigravity /teamwork for multi-agent scientific and engineering tasks

Details

Google Research has publicly released Antigravity /teamwork, a multi-agent system designed to tackle unsolved problems in mathematics, chemistry, and systems engineering by collaborating with human experts.
The system addresses LLM limitations in generation predictability and context dispersion by using parallel search with falsifiers (Lean 4, test suites) and splitting context among specialized agents with clean-context critics.
In mathematics, /teamwork proved the Courtade–Kumar conjecture (formalized in 50 million lines of Lean 4) and the Weil conjectures (formalized in 1.8 million lines of Lean 4), while also solving open problems in adaptive data analysis, quantum computing, and integer multiplication.
For software engineering, the tool rewrote the Zircon microkernel and libgav1 video decoder into memory-safe Rust, achieving 20-30% speedups over C++ and hand-optimized assembly, and optimized TPU kernels with a 4.09x geometric-mean speedup on JAXBench.

Key point