Google has highlighted a new use case for its Antigravity platform: autonomous teams of agents powered by Gemini 3.7 Flash working on math and engineering problems.
According to Google, the pairing has been used to tackle open math problems, build CPU emulators and optimize open-source software. The announcement is notable less for a single end product than for the operating model it describes: multiple agents working as a team on technical tasks that can require exploration, implementation and iteration.
What changed
Gemini 3.7 Flash is being positioned as the model behind autonomous agent teams in Antigravity. Google’s examples span three technically demanding categories:
- work on open mathematics problems;
- construction of CPU emulators; and
- optimization of open-source software.
The common thread is that these are not framed as simple question-answering tasks. They involve building, testing or improving artifacts, and potentially dividing work across a group of agents.
Why it matters for builders
For engineering leaders and founders, the important development is the move from an individual coding assistant to a coordinated-agent setup. A team structure can, in principle, assign different parts of a problem to separate agents rather than relying on one long interaction with a single model.
Google’s examples suggest where it sees this approach fitting: problems with enough depth to benefit from parallel technical work. A CPU emulator, for example, is a software-building task; optimizing an open-source project implies working against an existing codebase and its constraints. Open math problems represent a further step into research-oriented work.
That does not mean autonomous teams eliminate the need for technical review. The source describes what the system has been used to solve or build, but does not provide details on evaluation methods, deployment controls or the level of human oversight involved. Teams considering similar systems should treat those questions as central operational requirements.
A company signal, not just a model update
The announcement also shows Google tying model capability to an agent environment rather than presenting Gemini 3.7 Flash solely as a standalone model. Antigravity is the layer in which the agents are organized into teams, while Gemini 3.7 Flash supplies the model capability.
That combination matters competitively. The value proposition is increasingly about whether a company can offer a workable system for assigning, coordinating and assessing multi-step work—not merely a capable model for a single prompt.
For operators, the practical question is whether agent teams can produce useful technical output reliably enough to fit existing engineering and research workflows. The strongest early opportunities are likely to be bounded projects with concrete artifacts to inspect: code, emulator behavior, optimization results or mathematical reasoning.
What to watch next
Google’s next evidence will matter more than the examples alone. Watch for details on how Antigravity teams are configured, how their work is verified, what tasks they can complete without intervention, and whether the approach is made broadly available to developers.
Also watch whether Google provides reproducible results for the cited math, CPU-emulation and open-source optimization work. For companies evaluating multi-agent systems, the deciding metrics will be quality, reliability, review burden and the time required to turn agent output into production-ready work.