Google Research has reported new results from /teamwork, a feature of its Antigravity system that coordinates groups of AI agents, meaning several AI programs that cooperate on one job. In a blog post, Google reports that teams at Google and partner universities used /teamwork with Gemini 4 models to attack open problems in mathematics, science, chemistry, and engineering that had been stuck for years.
The main idea is to divide work among specialized agents. Each one receives only the files and tools needed for one step, which helps the system track details during tasks that run for days. Coordinator agents hand work to fresh successors, while critic and auditor agents check results independently. The system also routes each request to a workflow pattern that fits the task, such as /boost, a loop that solves a problem and then verifies the answer.
Results across science and engineering
In mathematics, the post says the system helped prove and verify the Courtade-Kumar conjecture and settle open questions in optimization, statistics, quantum computing, and algorithms. Many proofs were checked in Lean 4, software that turns a proof into code a computer can verify line by line, and one workflow called /teamwork-LEAP produced complete machine-checked formalizations of earlier results. In chemistry, pairing generator agents with independent verifier agents helped determine molecular structures from experimental measurements. A similar workflow turned 13 public data tables into a 50-task causal inference benchmark, using 29 agents over 11.5 hours.
In engineering, the team describes multi-day projects measured against objective tests. One effort finished a memory-safe rewrite in the Rust programming language of an existing codebase, ending up 20 to 30 percent faster than the original C++ and hand-optimized assembly. Other projects produced custom kernels, meaning small compute programs, for Google's TPU AI chips, and networking work connected to Snap, Google's microkernel-style host networking system. Inside Google, related workflows found memory savings and cut TPU workload interruptions.
The results matter because they suggest that groups of AI agents, restrained by independent checking, can contribute to real research and large engineering projects rather than only short demonstrations. Google says outside scientists and developers can already use Gemini 4 with /teamwork and /boost in Antigravity, with more features promised in the coming weeks.