Google Research releases Antigravity teamwork and Gemini 4 for scientific problem solving

According to Google Research:

Background and Design

Google Research details the design and applications of Antigravity's teamwork multiagent systems, built on top of various Gemini models including Gemini 4 Argon, Carbon, and Hydrogen. Frontier research requires navigating vast literature, building computational pipelines, and verifying every step. To address the unreliability of single model generations and the loss of context in multi-day tasks, teamwork uses parallel exploration paired with dedicated falsifiers like Lean 4 proof checkers and execution simulators, while partitioning context across specialized agents.

Scientific and Engineering Results

According to Google, teams across Google and academic institutions used these multi-agent workflows to solve long-standing open problems. In mathematics and theoretical computer science, researchers proved the 2013 Courtade–Kumar conjecture, resolved a 2014 open question in adaptive data analysis, and formalized complex proofs in Lean 4, including the Weil conjectures and parts of the Navier–Stokes equations. In systems software and hardware engineering, the tools were used to build the FarArray low-latency kernel-bypass system, rewrite large codebases like the Zircon microkernel and libgav1 into safe Rust, and optimize TPU kernels, yielding significant performance gains over manual implementations.

Availability

Google states that these features, along with the /boost command and Deep Review document evaluation tools, are currently available in Antigravity. Additionally, Google has opened applications providing 10,000 academic researchers with six months of free access to their highest usage tier.