Alphabet’s Google says Gemini-powered “Teamwork” agents solved open math, built a CPU simulator, and improved core open-source libraries

Google is expanding its Antigravity platform for multi-agent work, saying it paired its Gemini 3.7 Flash model with a “Teamwork” orchestration layer to accelerate progress on long-horizon research and engineering tasks.

In a post describing recent updates, Google frames Teamwork as a framework that lets autonomous AI agent teams collaborate, critique each other’s output, and iterate over hours or days when problems are too complex for a single pass. Google says the Gemini 3.7 Flash pairing tightened the loop from research ideation to verifiable results across multiple domains.

On the research side, Google says its agent teams solved seven open problems across top academic venues, listing areas that include Knuth’s Cycles Conjecture, sparse convex optimization, provable quantization for large language models, and prefix-matrix factorizations. For the Cycles Conjecture, Google says the result was verified in Lean using proofs totaling more than 40 pages.

Google also provided a benchmark figure for theoretical computer science work, saying the system achieved 71% on TCSBench. The company does not break down which sub-tasks drove the score or whether the evaluation set mirrors any training or prompt templates, but it presents the figure as evidence that the multi-agent approach can reach performance levels in reasoning-heavy settings.

In systems engineering, Google says it built a cycle-accurate, out-of-order RISC-V CPU simulator from scratch. Google’s stated milestone is that the simulator boots the xv6 operating system, runs through to a shell, and matches hardware ground truth with a 0.71% cycle-alignment error.

Cycle-alignment error is a specific way of describing how closely a simulator’s timing tracks a real processor’s cycle behavior. In practical terms, Google’s claim suggests its agents were not only able to produce functional code, but also to tune behaviors that align with expected timing, an area where many emulators can be “close but not exact.”

The update also highlights software engineering outcomes, with Google saying it landed performance optimizations upstream in core libraries. It cited Eigen, a popular C++ linear algebra library, saying it added “SIMD fast-paths” to improve performance on vectorized CPU operations.

Google also cited ParlayHash, describing changes it says doubled insert throughput and reduced memory usage by 25%. Upstreaming such improvements matters because it turns experimental work into broadly available code that other developers can use without integrating a separate proprietary system.

The broader business angle is that Alphabet’s research units are increasingly tying model performance to tools and workflows, not just standalone answers. A multi-agent orchestration layer like Teamwork is designed to make model outputs more durable by adding stages such as peer critique and repeated refinement, while “cycle-accurate” and “upstream” deliverables reflect goals that are closer to traditional engineering metrics than to conversational quality.

Still, Google’s post does not provide details that readers might expect for independent verification, such as the specific problem statements, hyperparameters, compute budgets, or how many agent iterations were required per task. It also does not clarify how the team prevented leakage from benchmarks or ensured that “open problems” were truly unsolved prior to the agent work. As a result, the claims are best read as internal performance and development milestones reported by Google, rather than as a complete third-party audit.

What to watch next is whether Google expands the update with reproducible artifacts, such as open benchmarks, released code, or standardized evaluation procedures for the Teamwork framework. If the company continues to connect Gemini-driven agent teams to measurable engineering outcomes, it could influence how developers adopt AI tooling, especially for workloads that require correctness checks, long iteration cycles, and integration into production-grade software stacks.