Google published a report on August 31 describing work by teams of AI agents in Antigravity, using Gemini 3.7 Flash. The company's summary grouped its claims into mathematics, systems engineering, and open-source software.
The reported projects included:
- Mathematical work included claims about open problems and proofs checked with Lean.
- A systems project produced a RISC-V processor simulator that could boot the xv6 operating system.
- Software work included reported optimizations to the Eigen and ParlayHash libraries.
The categories also call for different evidence. A proof should be examined with its formal statement and checking procedure. A simulator needs comparison against a reference behavior. A software optimization needs reproducible measurements and a look at what the upstream project accepted.
Google's Teamwork documentation describes separate research, implementation and verification roles, including review and stress-testing before milestones are approved. It also describes isolated project directories and assigned file ownership. The design is intended to let another process reject a proposed result and reduce the risk of agents overwriting each other's files.
Those workflow controls and the reported benchmark outcomes answer different questions. The controls describe how work is organized; a proof check or a reproducible performance result establishes whether a particular output succeeded.
The workflow described agents collaborating, criticizing results, and iterating over hours or days.
