What changed
Google Research and academic partners have used Antigravity’s /teamwork multiagent systems to solve open problems in mathematics, chemistry, and systems engineering. These tools are now being made available externally. The work leverages Gemini 4 models and multi-agent workflows to produce new innovations across information theory, formal verifications, molecular structure discoveries, and kernel optimization.
What this means in practice
The /teamwork system enhances human-AI collaboration by addressing two key limitations of current language models. First, it bridges the gap between a single attempt and finding the correct answer by pairing parallel exploration with dedicated falsifiers, such as Lean 4 proof checkers or software test suites. Second, it prevents AI models from losing track of details over multi-day tasks by partitioning context so each specialized agent receives only the exact files and tools needed for a specific step.
The system expresses workflows as declarative self-improving natural-language patterns and dynamically routes user prompts to the best-suited pattern. These patterns include distributed coding, iterative problem-solving, and specialized research workflows like long proof and document review. Researchers and developers can now apply these tools to their own work, with many features already available in Antigravity.
Limitations and availability
While many features are available in Antigravity, additional updates are expected in the coming weeks to provide all features to external users. The tools are being made available alongside /boost and other agents. The supplied announcement does not specify the exact release timeline or availability details for all features. The results include proofs and algorithms verified by researchers, with some examples involving Lean 4 formalizations and collaborations with institutions like CUHK, USC, and CMU.
