Google Research has described how its Antigravity /teamwork multi-agent workflow is being used with human specialists on problems in mathematics, science and software engineering. In its November 13 account, the company says the work combined several Gemini models with expert direction and tools that check candidate results.
The report is a description of Google and partner collaborations, not an independent evaluation of every result. Google says human experts framed problems and supplied constraints while agents explored alternatives, wrote computational tools and assisted with verification.
How the workflow divides work
According to Google, /teamwork partitions tasks among specialized agents, with coordinators passing work to fresh successors and reviewers examining results separately. It describes using checkers such as Lean 4, software tests, simulations and hardware counters to reject candidate solutions that fail specific checks.
That separation can make a long investigation easier to manage, but a successful automated check proves only what the check covers. A proof checker can validate a formally stated result without establishing that the formal statement captures every real-world assumption. Likewise, passing a software test suite does not prove a program is free of errors.
Interpreting the claims
Google says its collaborations produced advances across several research and engineering areas and that the tools are available externally through Antigravity. Readers should follow the underlying papers and project documentation for each claim, including author contributions, review status and access conditions; one company blog cannot establish the validity of every reported result.
The broader lesson is that agent orchestration and verification are different parts of a workflow. Independent replication, careful problem formulation and human review remain necessary, especially when a result will be relied on in research or production. AiBrainWorX has not independently reproduced the work.
Original reporting; primary sources reviewed November 14, 2026. No independent hands-on test claimed.
