AI

Antigravity /teamwork: Google multiagent AI systems for science

Google shares results from Antigravity's /teamwork multiagent systems powered by Gemini 4, solving open problems in math and engineering.

·3 min read
Antigravity /teamwork: Google multiagent AI systems for science

Google has shared recent results from Antigravity's /teamwork multiagent systems, which enhance human-AI collaboration to solve open problems across mathematics, chemistry, systems software, and machine learning. Developed in collaboration with human experts, these agentic tools are now being made available to external users.

  • Google and academic partners tackled open problems in math, science, and systems engineering that remained unsolved for years using Gemini 4 models and Antigravity's /teamwork.
  • The system bridges the gap between single generation attempts and correct answers by pairing parallel exploration with dedicated falsifiers like Lean 4 and execution simulators.
  • /teamwork partitions context by design: specialized agents receive only the exact files and tools they need, preventing attention dispersion across long tasks.
  • Innovations were produced in information theory, formal verification, molecular structures, and kernel optimization.

Architecture and Design of Multiagent Systems Frontier research requires navigating vast literature, building custom computational pipelines, and verifying every step. Progress comes from an interactive loop where researchers frame questions while long-horizon agents write custom solvers and check logic chains. /teamwork's design is motivated by two limits of current language models: unpredictable generation and loss of detail over multi-day tasks due to growing context windows. To bridge the gap, the system pairs parallel exploration with dedicated falsifiers, such as the Lean 4 proof checker, test suites, or hardware counters, to automatically verify candidate solutions. Additionally, /teamwork partitions context by construction, ensuring each specialized agent receives only necessary files. Coordinators hand off tasks to fresh successors, while critics verify results from clean contexts. ## Breakthroughs in Mathematics and Theoretical Computer Science Researchers from Google, CUHK, USC, and CMU used /teamwork's Long Proof pattern to prove information theory's 2013 Courtade–Kumar conjecture in full generality. Gemini models partitioned the parameter space and developed analytical bounds, formalizing the entire proof in Lean 4 to produce 50 million lines of code. This confirmed that compressing input to a single bit preserves no more information than a dictator function through a noisy channel. Researchers also posed open problems directly to Gemini 4 and multi-agent harnesses, receiving complete algorithms and proofs. In quantum computing, a new construction settled a 2014 open question by finding Boolean functions needing strictly more than n/2 quantum queries at constant error. Breakthroughs were also achieved in Boolean matrix multiplication, high-dimensional geometry, and dynamic graph streaming algorithms. ## Implications for Researchers and Developers Making these agentic tools available in Antigravity allows researchers and developers to apply advanced multi-agent workflows to their own work. By expressing workflows as declarative natural-language patterns and dynamically routing prompts, Antigravity enables customized specifications that dramatically accelerate discovery loops in science and engineering.

Source: Hacker News: Front Page

Related articles