Google Research has described a multi-agent system intended to help researchers and engineers pursue problems that require long searches, repeated testing and formal verification. The company says the system, called Antigravity /teamwork, divides work among specialized agents and connects their output to checkers that can reject incorrect candidates.
The design addresses two weaknesses Google identifies in current language models. A single model attempt can be unreliable, while a long-running conversation can accumulate so much context that important details are lost. In /teamwork, multiple agents explore possibilities in parallel, but dedicated verification tools decide which results survive. Depending on the task, those tools can include the Lean 4 proof checker, existing software test suites, execution simulators, web search or hardware performance counters.
Context is also separated deliberately. Each agent receives the files and tools needed for a specific stage, coordinators pass work to fresh agents, and critics review results without inheriting the producer's full reasoning trail. Google says workflows are expressed as natural-language patterns and routed according to the problem. A related /boost mode uses a sequential solve-and-check loop.
Google presented the software alongside a broad set of research claims. It said teams at Google and partner universities used Gemini 4 models and multi-agent workflows on questions in information theory, quantum computing, graph algorithms, formal verification and systems optimization. Among the examples was work on the Courtade-Kumar conjecture, where researchers and Gemini models reportedly developed bounds, partitioned a parameter space and formalized a proof in Lean 4.
Those results are described by Google and should be read as the company's account of collaborations that still depend on expert review. The announcement repeatedly frames the human role as defining questions and constraints, examining outputs and preparing results for publication. The automated component searches larger spaces, builds solvers and checks candidate solutions rather than replacing the researchers who determine whether a result is meaningful.
The system is also aimed at engineering tasks. Google says teams used agent workflows for kernel optimization and production infrastructure work, where unmodified tests and hardware counters can provide concrete acceptance criteria. That model is particularly suited to code, because an agent's proposal can be compiled, executed and measured.
Google says many of the described capabilities are available through Antigravity, with further external access planned. The announcement does not establish that every research problem can be treated this way. It instead outlines an architecture for cases where a task can be divided cleanly and candidate answers can be tested against formal or empirical checks.
