Antigravity's /teamwork and Gemini 4: Tackling Research Problems That Stayed Open for Years

type

Post

status

Published

date

Nov 14, 2026

slug

summary

Antigravity's /teamwork and Gemini 4: Tackling Research Problems That Stayed Open for Years

From open problems to verified results

Over the past several weeks, teams across Google and academic partner institutions have gone after open problems in mathematics, science, and complex systems engineering — questions that had resisted resolution for years. Their shared toolkit combined three Gemini 4 models, Gemini 4 Argon, Gemini 4 Carbon, and Gemini 4 Hydrogen, with multi-agent workflows such as Antigravity's /teamwork, which itself builds on various Gemini models.

The collaborations produced new results across several fronts: information theory, formal verification, molecular structure discoveries, and kernel optimization. Several of the same techniques also translated into production infrastructure gains at Google. To let other researchers and developers apply the same methods to their own work, these agentic tools are being made available in Antigravity's /teamwork, alongside /boost and other agents.

What frontier research actually demands

Frontier research and large-scale systems engineering share a demanding working pattern. Both require navigating vast bodies of literature, assembling custom computational pipelines, testing designs against physical measurements, and verifying every step along the way.

Progress in this setting rarely comes from a single model call. It emerges from an interactive discovery loop: researchers frame the questions and set the theoretical or physical constraints, while long-horizon agents write custom solvers, explore large search spaces, test candidate designs, and check long chains of logic. /teamwork was built to make that loop practical.

Two model limits that shaped the design

Two known limitations of current language models motivated /teamwork's architecture.

The first is unpredictability. Model generation can be erratic, leaving a gap between any single attempt and the correct answer. To close that gap, /teamwork pairs parallel exploration with dedicated falsifiers — components whose job is to knock down weak candidates automatically. These falsifiers include the Lean 4 proof checker, unmodified software test suites, execution simulators, web search, and hardware counters. Rather than trusting a plausible-looking answer, the system verifies which candidate solutions actually hold up.

The second limit is drift on long tasks. Over efforts spanning multiple days, models tend to lose track of details as attention disperses across ever-growing context windows. /teamwork counters this by partitioning context by construction: every specialized agent is given just the specific files and tools required for its own step, and nothing more. Coordinators hand tasks to fresh successors, while critics and auditors challenge and verify results from clean contexts — deliberately never exposed to the original producer's line of reasoning, so their judgment is not anchored by it.

Declarative, self-improving patterns

/teamwork expresses these workflows as declarative, self-improving natural-language patterns and dynamically routes each user prompt to the pattern best suited to the task.

The pattern library spans a wide range. One end covers distributed coding and iterative problem-solving, also packaged on its own as /boost — a sequential solve-and-verify loop refined from prior trajectories. The other end covers specialized research workflows such as long proof construction and document review. All of these are available in Antigravity today.

Beyond the built-in patterns, teams can design custom specifications from past run transcripts, routing work across different harness patterns and dynamically improving the agentic harness itself. Google notes that while many of these features are already in Antigravity, further updates will arrive in the coming weeks to bring all of the features to external users.

A full proof of the Courtade–Kumar conjecture

In mathematics and theoretical computer science, researchers paired Gemini 4 models with /teamwork harnesses to collaborate on large-scale proofs and to generate complete algorithms and proofs directly from open problem statements.

The clearest example concerns the 2013 Courtade–Kumar conjecture in information theory. Researchers from Google, CUHK, USC, and CMU used /teamwork's Long Proof pattern — which implements Stellar Colosseum inside Antigravity — to prove the conjecture in full generality.

The result confirms that when a random bit string passes through a noisy channel, no way of compressing the input down to a single bit preserves more information than a *dictator function* — one that simply copies a single input bit.

The path to the proof was genuinely collaborative:

1. Gemini first surfaced counterexamples to a generalized version of the conjecture from the CUHK team's earlier paper.

2. The CUHK team responded with a new candidate function that reduced the conjecture to a 14-variable optimization problem, together with a key analytical lower bound.

3. Gemini 4 Argon and earlier Gemini models then partitioned the parameter space and developed most of the new analytical bounds and interval-arithmetic certificates.

4. The models went on to extend the proof from balanced to unbalanced functions without human input.

5. Finally, the team formalized the entire proof in Lean 4, producing 50 million lines of code.

Open problems beyond information theory

The same approach was applied more widely. Researchers posed open problems in optimization, statistics, quantum computing, and algorithms directly to Gemini 4 and to multi-agent harnesses such as Stellar Colosseum and /teamwork, receiving end-to-end algorithms and proofs, often with Lean 4 involved.

Tools researchers can pick up now

Alongside the research results, the practical takeaway is access. The agentic tools used in these collaborations are available in Antigravity's /teamwork, alongside /boost and other agents, giving researchers and developers a way to apply the same verify-as-you-go workflows to their own problems. Google has also said that additional features will reach external users on Antigravity in the coming weeks. For anyone working on problems that demand long reasoning chains and rigorous verification — precisely the conditions /teamwork was designed for — the workflows behind these results can now be tried directly in Antigravity.

Source: Google Research, https://research.google/blog/advancing-science-engineering-with-antigravitys-teamwork-feature/