Inside Antigravity's /teamwork: How Gemini 4 Multi-Agent Workflows Are Advancing Science and Engineering

type

Post

status

Published

date

Nov 14, 2026

slug

summary

Inside Antigravity's /teamwork: How Gemini 4 Multi-Agent Workflows Are Advancing Science and Engineering

Over the past several weeks, teams across Google and its academic partner institutions have taken on open problems in mathematics, science, and complex systems engineering—problems that had remained unresolved for years. Their shared toolkit paired Gemini 4 Argon, Gemini 4 Carbon, and Gemini 4 Hydrogen with multi-agent workflows such as Antigravity's /teamwork, which builds on a range of Gemini models.

The collaborations produced new results across information theory, formal verification, molecular structure discovery, and kernel optimization, along with several production infrastructure gains at Google. The agentic tools behind this work are available in Antigravity's /teamwork, alongside /boost and other agents, so researchers and developers can apply the same methods to their own problems.

What the collaborations delivered

New innovations in information theory, anchored by a full proof of the 2013 Courtade–Kumar conjecture (detailed below)

Progress in formal verification, with large proofs machine-checked in Lean 4

New molecular structure discoveries

Kernel optimization improvements

Several production infrastructure gains at Google

Why frontier research strains current models

Frontier research and systems engineering demand more than one-shot answers. The work requires navigating vast literatures, building custom computational pipelines, testing designs against physical measurements, and verifying every step of the process.

Progress, as Google describes it, comes from an interactive discovery loop: researchers frame the questions and define 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.

Two limits, two design answers

/teamwork's architecture responds directly to two known weaknesses of today's language models.

Unpredictable generation. Model output can be erratic, leaving a wide gap between a single attempt and the correct answer. To close that gap, the system pairs parallel exploration with dedicated falsifiers—the Lean 4 proof checker, unmodified software test suites, execution simulators, web search, and hardware counters—which automatically verify which candidate solutions actually work.

Attention drift on long tasks. Over efforts spanning multiple days, models tend to lose track of details as attention disperses across growing context windows. /teamwork counters this by partitioning context by construction: each specialized agent receives only the exact files and tools it needs for its specific step. Coordinators hand tasks off to fresh successors, while critics and auditors challenge and verify results from clean contexts, without seeing the original producer's reasoning. The goal is to keep multi-day investigations coherent even as the total context grows far beyond what a single window can comfortably hold.

Declarative, self-improving workflow patterns

These workflows are expressed as declarative, self-improving natural-language patterns, and the system dynamically routes each user prompt to the pattern best suited to the task.

The patterns cover a wide spectrum. They range from distributed coding and iterative problem-solving—packaged standalone as /boost, a sequential solve-and-verify loop refined from prior trajectories—to specialized research workflows such as long proof construction and document review. All of these are available in Antigravity today.

Teams can go further by designing custom specifications from past run transcripts, routing across different harness patterns and dynamically improving the agentic harness. Google notes that while many of these features are already live in Antigravity, additional updates will bring the full feature set to external users in the coming weeks.

Case study: proving the Courtade–Kumar conjecture

The clearest example comes from mathematics and theoretical computer science, where researchers paired Gemini 4 and /teamwork harnesses to collaborate on large-scale proofs and to generate complete algorithms and proofs directly from open problem statements.

A team spanning Google, CUHK, USC, and CMU used /teamwork's Long Proof pattern—which implements Stellar Colosseum in Antigravity—to prove information theory's 2013 Courtade–Kumar conjecture in full generality, settling a question that had been open for over a decade.

The conjecture concerns what survives a noisy channel: when a random bit string passes through one, no compression of the input to a single bit preserves more information than a "dictator function," which simply copies a single input bit. In plain terms, the best way to squeeze a random bit string down to one bit through a noisy channel is to copy one of the input bits.

The proof came together through an iterative human–AI exchange. Gemini first found counterexamples to a generalized version of the conjecture from the CUHK team's earlier paper. The CUHK team then proposed a new candidate function that reduced the conjecture to a 14-variable optimization problem, along with 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. The system also extended the proof from balanced to unbalanced functions without human input.

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

Beyond information theory

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

This direct-to-model workflow—stating an open problem and receiving a complete algorithm and proof in return—complements the more interactive, multi-session style of collaboration seen in the Courtade–Kumar effort.

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