How Antigravity's /teamwork and Gemini 4 Models Are Accelerating Frontier Research in Math and Engineering

type

Post

status

Published

date

Nov 14, 2026

slug

summary

How Antigravity's /teamwork and Gemini 4 Models Are Accelerating Frontier Research in Math and Engineering

Recent collaborative efforts between Google and academic institutions have demonstrated that advanced agentic tools can effectively tackle open problems that have remained unsolved for years. By leveraging a combination of the latest Gemini 4 models—specifically Argon, Carbon, and Hydrogen—alongside multi-agent workflows powered by Antigravity's /teamwork feature, researchers have achieved breakthroughs in information theory, formal verification, and molecular structure discovery.

The Challenge of Long-Horizon Reasoning

Traditional AI models often struggle with two key limitations when handling complex, multi-day research tasks. First, generation is unpredictable; a single attempt rarely yields the correct answer without verification. Second, models lose track of details as context windows grow, causing attention to disperse.

To address this, /teamwork employs a design that pairs parallel exploration with dedicated "falsifiers." These include proof checkers like Lean 4, unmodified software test suites, and execution simulators that automatically verify candidate solutions. Furthermore, the system partitions context by construction. Each specialized agent receives only the specific files and tools needed for its step. Coordinators hand off tasks to fresh successors, while critics and auditors verify results from clean contexts, preventing contamination from the original producer's reasoning. This allows for the execution of declarative, self-improving natural-language patterns that dynamically route user prompts to the best-suited workflow.

Breakthroughs in Mathematics and Theoretical Computer Science

One of the most significant applications of these tools is in mathematics and theoretical computer science. Researchers have used Gemini 4 and /teamwork harnesses to generate complete algorithms and proofs directly from open problem statements.

Proving the Courtade–Kumar Conjecture

A team comprising researchers from Google, CUHK, USC, and CMU utilized /teamwork's Long Proof pattern to resolve the 2013 Courtade–Kumar conjecture in full generality. This information theory conjecture posits 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 simply copies a single input bit).

The process began with Gemini identifying counterexamples to a generalized conjecture proposed in an earlier CUHK paper. The CUHK team then proposed a new candidate function, reducing the conjecture to a 14-variable optimization problem and establishing a key analytical lower bound. Gemini 4 Argon and earlier Gemini models partitioned the parameter space and developed most of the new analytical bounds and interval-arithmetic certificates. Notably, the models extended the proof from balanced to unbalanced functions without human intervention. The entire proof was formalized in Lean 4, resulting in 50 million lines of code.

Addressing Open Questions in Optimization and Quantum Computing

Beyond information theory, researchers posed open problems in optimization, statistics, and quantum computing directly to Gemini 4 and multi-agent harnesses. The systems produced end-to-end algorithms and proofs, frequently utilizing Lean 4 for formal verification. This approach demonstrates the potential of agentic systems to navigate vast literature and complex logical structures, moving beyond simple code generation to high-level theoretical research.

Making Agentic Tools Accessible

The infrastructure supporting these results is built on multi-agent workflows that interact with physical measurements and custom computational pipelines. While many of these features are currently available within Antigravity, the goal is to provide these capabilities to external users in the coming weeks. The integration of /boost, a sequential solve-and-verify loop, alongside /teamwork, allows teams to design custom specifications based on past run transcripts. This dynamic routing across different harness patterns ensures that the agentic framework improves over time, bridging the gap between human-framed questions and machine-verified solutions.

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