← Back to BLACKWIRE EMBER BUREAU AI RESEARCH BREAKTHROUGH Diagram of Antigravity /teamwork multi‑agent workflow with Gemini models and verification loops

The /teamwork architecture routes prompts through specialized agents, coordinators and auditors to produce verified results.

GOOGLE'S ANTIGRAVITY /TEAMWORK SOLVES LONG-OPEN MATH CONJECTURES

*Google Research unveiled multi‑agent breakthroughs that turned years‑old theory problems into verified code. The results span information theory, adaptive data analysis and quantum computing, and the tools are slated for external release.*

By EMBER Bureau - BLACKWIRE  |  November 14, 2026, 18:00 CET  |  Antigravity, /teamwork, Gemini 4, Courtade–Kumar conjecture, multi‑agent AI

Google Research unveiled a suite of multi‑agent tools under the Antigravity /teamwork banner on November 13, 2026. Vahab Mirrokni, Google Fellow and VP of Google Research, detailed how Gemini 4 Argon, Carbon and Hydrogen models, combined with /teamwork workflows, have cracked problems that lingered for years. The effort spanned mathematics, chemistry, systems software and ML engineering, delivering new information‑theory proofs, formal verifications and kernel optimizations. All work was done alongside human experts, and the resulting agentic tools are slated for external release. The announcement signals a shift from isolated AI experiments to production‑grade, verifiable AI‑human collaborations that can scale across scientific domains.

Did /teamwork actually prove the Courtade–Kumar conjecture?

In November 2026, researchers from Google, CUHK, USC and CMU announced a full proof of the 2013 Courtade–Kumar conjecture using the /teamwork Long Proof pattern. Gemini 4 first generated counterexamples to a broader claim, prompting the CUHK team to narrow the problem to a 14‑variable optimization. Gemini 4 Argon and earlier models partitioned the space, derived new analytical bounds and extended the proof to unbalanced functions without human input. The entire argument was formalized in Lean 4, producing 50 million lines of code. The result confirms that a noisy channel cannot preserve more information through any compression than a simple dictator function that copies a single input bit.

How does /teamwork overcome language‑model limits?

The system tackles two core limits of current language models. First, unpredictable generation leaves a gap between a single attempt and the correct answer. /teamwork bridges this by pairing parallel exploration with dedicated falsifiers—Lean 4 proof checker, unmodified test suites, simulators, web search and hardware counters—to automatically verify candidates. Second, attention dilutes over multi‑day tasks. /teamwork partitions context by design: each specialized agent receives only the files and tools needed for its step. Coordinators hand off tasks to fresh successors, while critics and auditors verify results in clean contexts, preventing drift and ensuring reproducibility across long‑horizon reasoning.

The result confirms that when a random bit string passes through a noisy channel, no compression of the input to a single bit preserves more information than a "dictator function", which just copies a single input bit.

What new algorithms emerged from the multi‑agent runs?

Beyond the Courtade–Kumar proof, /teamwork delivered end‑to‑end algorithms in adaptive data analysis and quantum computing. A Gemini 4 Argon harness independently proved and Lean‑verified near‑optimal sample bounds for adaptively chosen statistical queries, closing a gap open since 2014. Those bounds enable the first nearly space‑optimal algorithm for data streams with insertions and deletions that remains accurate against adaptive adversaries. In quantum computing, the same multi‑agent pipeline settled a 2014 open question by constructing the first n‑bit Boolean functions needing a specific resource, demonstrating the system’s reach into previously intractable quantum design problems.

What is Google’s plan for external access to /teamwork?

Google announced that the agentic tools—including /teamwork, /boost and related Gemini 4 models—will be made available externally through the Antigravity platform. Researchers and developers can apply the multi‑agent workflows to their own problems via the Antigravity researcher access program. The rollout will follow the current internal successes, with updates promised in the coming weeks. By exposing the declarative self‑improving patterns and dynamic routing of user prompts, Google aims to let external teams design custom specifications, reuse past run transcripts and benefit from the same verification loops that powered the recent mathematical breakthroughs.

Frequently asked questions

What is the Courtade–Kumar conjecture?

It posits that when a random bit string passes through a noisy channel, no compression to a single bit preserves more information than a "dictator function" that copies a single input bit.

How many lines of Lean 4 code were generated for the proof?

The formalized proof in Lean 4 produced 50 million lines of code.

Which institutions collaborated on the Courtade–Kumar proof?

Researchers from Google, CUHK, USC and CMU worked together using /teamwork's Long Proof pattern.

Google’s public rollout of /teamwork marks a decisive moment for AI‑augmented research. By exposing verified, self‑improving agents to the broader scientific community, the company promises to accelerate discovery cycles that previously stalled for years. The next wave of external projects will test whether the internal breakthroughs can survive the scrutiny of independent labs. If they do, the balance of power in high‑stakes scientific research could tilt sharply toward AI‑driven collaboration.

Sources: Hacker News – Advancing science and engineering with Antigravity’s /teamwork