Research review / research

Google's /teamwork Puts Checkers, Not Bigger Prompts, at the Centre of Multi-Agent Research Work

Google Research described results from Antigravity's /teamwork on November 13: a proof of the 2013 Courtade-Kumar conjecture formalized in 50 million lines of Lean, an autonomous Weil conjectures formalization, and 95% top-1 accuracy on molecular structure puzzles with 30 agents. The design rests on two choices: independent falsifiers and agents that each see only what they need.

By Sandeep Belgavi · Published 2026-11-14

On November 13, 2026, Vahab Mirrokni of Google Research published a summary of recent work done with /teamwork, the multi-agent mode in Google's Antigravity development platform, running on Gemini 4 models. The post says "all of this work is done in collaboration with human experts" and that the tools are now available externally.

Two design choices

The post names "two limits of current language models" behind the design. The first is that a single attempt is unreliable, so the system pairs "parallel exploration with dedicated falsifiers (e.g., the Lean 4 proof checker, unmodified software test suites, execution simulators, web search, or hardware counters)" that decide which candidates work. The second is that models "often lose track of details over multi-day tasks because attention disperses across growing context windows." /teamwork therefore "partitions context by construction: each specialized agent receives only the exact files and tools it needs for a specific step." Coordinators hand work to fresh successors, and "critics and auditors challenge and verify results from clean contexts without the original producer's reasoning."

What Google reports

  • Information theory. Researchers from Google, CUHK, USC and CMU proved the 2013 Courtade-Kumar conjecture "in full generality". Humans reduced it to a 14-variable optimization problem, the models produced most of the new bounds and certificates, and "the team formalized the entire proof in Lean 4, producing 50 million lines of code."
  • Formalization. A framework called LEAP "fully autonomously produced an end-to-end, machine-checked formalization of the Weil conjectures", spanning "about 1.8 million lines of Lean, produced across 28,800 subagent invocations."
  • Chemistry. "By organizing 30 specialized agents into a self-correcting loop" with a spectral simulator, the system reached "95% top-1 exact match and 100% top-3 accuracy on complex benchmark molecules, nearly doubling single-agent performance."
  • Benchmarks. "29 agents running on Argon over 11.5 hours turned 13 public tables into a 50-task causal inference benchmark", with reviewer and auditor agents that "caught shortcuts, flagged wording hints, and vetoed hallucinated scripts".
  • Kernels. A kernel search reports "a 4.09x geometric-mean speedup over reference implementations across 17 JAXBench kernels".
  • Review. Authors of 578 STOC 2027 submissions opted into Deep Review. In an ongoing survey of 70 authors, "90% of respondents rated the feedback mostly or almost all factual."

Why the design matters more than the list

Many of these results share one feature: an external checker that does not care how persuasive the model sounds. Lean either compiles or it does not. A test suite passes or fails. That is what makes 28,800 subagent calls usable, because each output can be accepted or rejected mechanically. Where no such checker exists, as with the review survey, the evidence is weaker: 70 self-selected respondents rating feedback as factual.

The context rule is the second lesson. Instead of one agent with an ever larger history, each step gets a fresh agent with a narrow set of files and tools, and the reviewer does not see the author's reasoning. That limits how far one early mistake can spread and makes the reviewer harder to persuade.

What remains uncertain. These are Google's own reports, many describing papers still in review. A 50-million-line Lean proof is machine-checked, but humans must still confirm that the formal statement matches the original problem, as the post says. The Go engine it describes sits "between KataGo at 16 and 32 visits per move", which is a modest level. Costs and compute per result are not given.

Analysis: At Belgavi's Book Store, Jenya's team could apply both rules to the agent that writes stock reconciliation scripts. Give the writing agent only the inventory schema and a sample file, run each script against a fixed set of known book orders with expected totals, and let a second agent review the script without seeing the first agent's notes. A script that fails the totals is rejected without discussion.

Sources and evidence

Google Research's November 13 /teamwork report credits two design choices: independent falsifiers such as Lean, test suites and simulators, and agents that see only the files and tools for one step. Reported results include a Courtade-Kumar proof formalized in 50 million lines of Lean, a 1.8 million-line Weil conjectures formalization and 95% top-1 accuracy on molecule structures. All are Google's own reports.

Views are my own and do not represent any employer, past or present.