Google says that pairing its Antigravity platform with the Gemini 3.7 Flash model has let autonomous AI agent teams crack seven open problems in math and theoretical computer science, including Knuth's Cycles Conjecture. The work ran through Teamwork, a framework in Antigravity that lets groups of AI agents collaborate, critique each other, and iterate over hours or days on complex tasks.
According to Google, the agents solved problems across top venues like FOCS and JMLR, verified Knuth's Cycles Conjecture in the Lean proof assistant with proofs spanning more than 40 pages, and also tackled sparse convex optimization, provable LLM quantization, and prefix-matrix factorizations. The same setup scored 71% on TCSBench, a benchmark for theoretical computer science.
Beyond math, the agents built a cycle-accurate, out-of-order RISC-V CPU simulator from the ground up that boots the xv6 operating system to a shell, with a cycle alignment error of just 0.71% against hardware ground truth. In open-source software, they landed performance improvements upstream in core libraries, including SIMD fast-paths in Eigen and a 2x insert throughput boost with a 25% memory reduction in ParlayHash.
Google announced the results in a blog post on August 31, 2026, and points to the Antigravity blog for more details. The announcement follows a series of recent Gemini model updates, including the rollout of Gemini 3.7 Flash and the API availability of Gemini Omni 1.1 Flash.
The company frames these wins as evidence that pairing a fast model with multi-agent orchestration can accelerate research and engineering work. It did not disclose whether the agents' solutions have been formally peer-reviewed or published beyond the venues named.
