Google says its Antigravity development platform is using teams of Gemini 3.7 Flash agents to tackle long-running math, systems and open-source software work. In an August 31 post on the Google Blog, the company described updates to Teamwork, a framework in which autonomous agents collaborate, critique one another and iterate over hours or days.
The post gives the framework a more concrete test than a short coding demo. Google says the teams solved seven open problems across venues including FOCS and JMLR, with one result, Knuth’s Cycles Conjecture, verified in Lean using proofs longer than 40 pages. It also reports a 71% score on TCSBench. These are results reported by Google, not an independent benchmark audit, and the announcement does not publish enough methodology to reproduce every claim from the article alone.
Teamwork is a different proposition from asking one chatbot to produce a larger answer. Google describes a group of agents that can divide work, review intermediate results and continue operating across a long horizon. That makes orchestration the product story. Neon Control’s WordPress 7.1 and AI agents coverage looked at how discovery and permission boundaries can make tool-using agents more practical. The site’s Gemini 3.7 Flash coverage provides model context, while the AI provider comparison tracks the difference between model claims and provider evidence. Antigravity applies the same concern to research and engineering tasks, where the hard part is coordinating many steps without losing track of verification.
From proofs to a CPU simulator
Google says the system also built a cycle-accurate, out-of-order RISC-V CPU simulator from scratch. The simulator booted the xv6 operating system to a shell, and Google reports a 0.71% cycle alignment error against hardware ground truth. That result is notable because it connects language-model work to a systems artifact with a measurable reference point. It still says nothing about whether the generated simulator is maintainable, portable or ready for use outside the reported experiment.
The same announcement points to changes in open-source libraries. Google says the agent teams landed performance optimizations in Eigen, including a SIMD fast path, and in ParlayHash. For ParlayHash, the company reports twice the insert throughput and a 25% reduction in memory use. The post does not frame those figures as universal gains across every processor or workload, so they should be read as results from the projects and tests Google selected.
Gemini 3.7 Flash is the model paired with the Teamwork updates. Google calls it an agent-first workhorse, but the new post focuses on the multi-agent harness rather than announcing a separate model release. The harness lets agents critique and iterate, while the long-running workflow gives them time to pursue a result that may require many failed approaches before a useful one emerges.
That shift changes what developers need to inspect. A single response can be judged as text; a multi-agent run needs evidence for task assignment, intermediate checks, tool use and the final artifact. The RISC-V result has a stated comparison target, while the mathematics and library claims still depend on Google’s own account. The next meaningful question is whether outside researchers can reproduce the work and whether the Teamwork framework exposes enough controls to make that evaluation practical.
Google’s announcement positions Antigravity as a research and engineering workspace, not an autonomous replacement for review. Its strongest promise is persistence: agents can keep working, challenge each other’s output and return a result tied to a concrete artifact. That is a more useful direction for agent software than simply claiming a higher benchmark score, provided the verification trail remains visible.
