Google Says Its /teamwork AI Agents Solved Long-Open Problems, Opens Them to Researchers

In a Google Research blog post published on Friday, Vahab Mirrokni, a Google Fellow and vice president at Google Research, described results from /teamwork, a multi-agent system in Antigravity. He wrote that the work was done with human experts at Google and partner universities and covered mathematics, chemistry, systems software and machine learning engineering.

For companies, the post is a concrete test of whether AI agents can take on long research and engineering tasks with checkable results. Google is backing that claim with free access for 10,000 academic researchers.

How the System Checks Its Work

Mirrokni said /teamwork is designed around two weaknesses of current language models. The first is that a single attempt from a model is unreliable. The system therefore runs many attempts in parallel and passes them to what Google calls falsifiers, such as the Lean 4 proof checker, unmodified software test suites, simulators and hardware counters, which confirm which candidate solutions actually work.

The second is that models lose track of detail over tasks lasting several days. To deal with this, /teamwork gives each specialised agent only the files and tools it needs for its step. Coordinators hand work on to fresh agents, and separate critic and auditor agents review results without seeing the original agent's reasoning. A simpler loop that solves a problem and then verifies the answer is packaged on its own as /boost.

Proofs, Molecules and Code

The best-known result is a proof of the Courtade-Kumar conjecture in information theory, posed in 2013. Researchers from Google, CUHK, USC and CMU proved it in full generality. According to the post, Gemini models first found counterexamples to a broader version of the conjecture, then developed most of the new analytical bounds and extended the proof to unbalanced functions without human input. The full proof was formalised in Lean 4 and runs to 50 million lines of code. Google also said its agents produced a machine-checked formalisation of the Weil conjectures, about 1.8 million lines of Lean generated across 28,800 subagent calls, and closed a gap in adaptive data analysis that had been open since 2014.

In chemistry, a system of 30 agents paired with a spectral simulator worked out molecular structures from NMR data. Google said it identified the exact structure as its first answer 95% of the time on complex benchmark molecules, and among its top three answers every time. That is close to double the accuracy of a single agent.

In engineering, agents rewrote the libgav1 AV1 video decoder in about 170,000 lines of Rust. Google said the rewrite matches the original bit for bit and, after optimisation, runs 20% to 30% faster than the C++ and hand-tuned assembly code on x86_64 and ARM chips. Inside Google, two agent-designed changes to its Borg cluster scheduler cut preemption of TPU workloads by 24% on some hardware types. A fleet-wide memory project saved more than the 1 PiB initially estimated.

All of these results come from Google's own account. The company said human researchers verified and edited them for publication, and many of the proofs were checked by machine in Lean 4.

Gemini 4 models with /teamwork and /boost are available in Antigravity now, Google said, along with Deep Review, a tool for checking manuscripts that includes two free calls a week. Authors of 578 submissions to the STOC 2027 computer science conference opted into Deep Review. Applications are open for 10,000 academic researchers to receive six months of free access at the highest usage tier. Google said a formal-proof workflow, a code-rewrite workflow and other specialised workflows will follow in the coming weeks.