Google published a batch of new results from Antigravity's multi-agent workflow /teamwork on its official research blog on November 13: over the past few weeks, Google teams and academic partners used the Gemini 4 model family together with /teamwork to settle several long-open problems in information theory, formal verification, molecular structure and kernel optimization. These agentic tools are now available to researchers and developers through Antigravity's /teamwork, no longer just an experiment inside Google.
Decades-old problems now have machine proofs
The eye-catching results are in mathematics. Working with researchers from the Chinese University of Hong Kong, the University of Southern California and Carnegie Mellon University, Google used /teamwork's Long Proof pattern to prove in full generality the Courtade–Kumar conjecture from 2013, formalizing the entire proof in Lean 4 — about 50 million lines of code. /teamwork-LEAP also independently produced an end-to-end, machine-checked formalization of the Weil conjectures, about 1.8 million lines of Lean code across 28,800 subagent invocations; a proof of Hilbert's tenth problem over general rings of integers and a proof of finite-time blowup for the 3D Navier–Stokes equations, open for about 90 years, also received machine-checked versions for the first time.
It works by parallel exploration plus independent falsification
/teamwork is designed around two weaknesses of long-horizon tasks. First, single attempts are unreliable, so it pairs parallel exploration with dedicated falsifiers — a Lean 4 proof checker, unmodified software test suites, execution simulators and hardware counters — that automatically discard wrong candidates. Second, long tasks lose context, so it partitions context by construction: each specialized agent receives only the files and tools its current step needs, coordinators hand tasks to fresh successors, and critics and auditors re-check results from clean contexts. In molecular structure determination, 30 specialized agents in a self-correcting loop, paired with a physics-informed spectral simulator, reached 95% top-1 exact match and 100% top-3 accuracy on complex benchmark molecules, nearly double single-agent performance.
Who can use it, and for what
Google says /teamwork is now available to researchers and developers through Antigravity, with more features arriving in the coming weeks. For research users, the most realistic use is not replacing researchers but handing long, grinding work — writing solvers, searching large spaces, checking every step — to a team of agents, while people frame the questions, set the constraints and verify the conclusions. The boundary is worth stating too: the results in the blog are Google's own account, and formal proofs are checked line by line by the Lean kernel, so the mathematical claims are more reliable than ordinary model output; what will really decide its value is whether ordinary researchers can reproduce the same workflow on their own problems.
