The Pulse
Google reports new proof using paid Antigravity Teamwork
Google says researchers used Antigravity’s multi-agent Teamwork workflow to prove the Courtade–Kumar conjecture and formalize the proof in Lean 4. The /teamwork command is available on paid plans, with further features promised for external users.

Researchers using Google’s Antigravity multi-agent workflow have proved a longstanding information theory conjecture, Google Research says. The collaboration formalized the proof in Lean 4, producing 50 million lines of code.
The /teamwork command is available on paid plans in the Antigravity app and command-line interface. In its November 13 research post, Google said further features used in the work would be made available to external users in the coming weeks.
A proof with human input
The Courtade–Kumar conjecture, posed in 2013, concerns how much information about a random bit string can be preserved when the string passes through a noisy channel and its input is compressed to one bit. The conjecture says no such compression retains more information than a function that copies a single input bit.
Researchers from Google, CUHK, USC and CMU worked on the problem with Teamwork’s Long Proof pattern. Gemini first found counterexamples to a generalized conjecture in an earlier paper by the CUHK team. The CUHK researchers then proposed a candidate function that reduced the problem to a 14-variable optimization problem, along with an analytical lower bound.
Google said Gemini 4 Argon and earlier Gemini models divided the parameter space and developed most of the new analytical bounds and interval-arithmetic certificates. The models also extended the proof from balanced to unbalanced functions without human input, according to the company.
How the agents divide work
Teamwork routes prompts to workflows suited to tasks such as distributed coding, proof work and document review. The system can explore candidates in parallel and use checks such as the Lean 4 proof checker, software test suites, execution simulators and hardware counters.
In Teamwork, specialized agents receive the files and tools needed for their assigned steps. Coordinators pass work to fresh agents, while critics and auditors examine results without the original producer’s reasoning. The available workflows include /boost, a separate sequential solve-and-verify loop.
Users first go through a scoping interview that sets requirements and checks before approving execution, according to Antigravity’s documentation. After the user approves the prompt, a project orchestrator coordinates the work through milestones without step-by-step prompts.
Other research and review results
Google also reported results in adaptive data analysis, where it said a multi-agent system using Gemini 4 Argon proved and Lean-verified near-optimal sample bounds for statistical queries selected using earlier answers. Another construction settled a 2014 quantum computing question, the company said.
The researchers said a separate framework, /teamwork-LEAP, autonomously produced a machine-checked formalization of the Weil conjectures across about 1.8 million lines of Lean and 28,800 subagent invocations. That was a formalization of existing mathematical results, rather than a claim to have newly proved the conjectures.
Google noted that human mathematicians must confirm a formal theorem statement matches the original problem; the Lean proof then compiles only if its logical steps hold.
The company also tested its Deep Review tool with authors preparing submissions for STOC 2027. Google said authors of 578 submissions, more than a third of the total, opted to receive feedback on their drafts.
In an ongoing survey of 70 authors, 90% rated the feedback mostly or almost entirely factual and 90% said they would use the tool again. More than 30% said it caught substantive proof bugs or mathematically significant typos, while 14% said it found technical gaps that took more than an hour to fix.
Sources
Companies in this story
Related stories
- Google Research tests wearable models for CKM health coaching
- Meta’s Muse can call US businesses for paid subscribers
- LiveGeo tests find AI agents struggle as geographic conditions change
- MatrAIx opens proposals for long-horizon personal agent benchmark
- Microsoft study finds Kimi K3 probe signals can arrive late