Google Research has published a long list of results from /teamwork, a multi-agent feature in its Antigravity coding environment, and the design matters more than any single result. Rather than trusting one model to get a hard task right, /teamwork runs many attempts in parallel and lets an outside checker decide which ones work. For developers using Antigravity with Gemini 4 Argon or Gemini 4 Carbon, it shows where Google thinks agent work is heading.
The post names two weaknesses the system is built around. A single model attempt is unpredictable, so /teamwork pairs parallel exploration with what Google calls falsifiers: the Lean 4 proof checker, unmodified test suites, simulators, web search or hardware counters. And models lose track of detail over multi-day tasks, so each agent sees only the files and tools needed for its step, coordinators hand work to fresh agents, and critics check results without seeing the original agent's reasoning. A simpler sequential version ships as /boost for everyday coding.
The claimed results are broad. Researchers say the system helped prove the 2013 Courtade-Kumar conjecture in information theory, formalised in Lean 4 as 50 million lines. A formalisation of the Weil conjectures spans about 1.8 million lines of Lean from 28,800 sub-agent calls. In software, a two-week campaign added about 800,000 lines of safe Rust to finish a port of the Zircon microkernel that passes its unit tests, and a Rust rewrite of Google's libgav1 video decoder ended up 20-30% faster than the C++ original. In chemistry, 30 agents reached 95% top-1 accuracy on molecular structure benchmarks. Ahead of STOC 2027, authors of 578 submissions used its Deep Review feature.
This is the same pattern that makes formal tools valuable elsewhere: the model proposes, a checker that cannot be argued with disposes. Developers who already run agents in Claude Code or Cursor will recognise the idea of separating the agent that writes from the agent that reviews. Google is opening the tools in Antigravity and offering 10,000 academic researchers six months of free access to its highest usage tier.
The caveats are significant. Almost every result is reported by Google and its collaborators in a blog post, and most of the maths claims await peer review. A Lean proof is only as good as the human check that its formal statement matches the original problem, which the post acknowledges. The compute cost of these campaigns is not disclosed, and some features described are not yet available to outside users.