Google says Gemini 4 agents solved open math and code problems

On 13 November 2026, Google Research shared results from /teamwork, a tool in its coding platform. The tool runs a group of AI agents, which work together on one task, on Gemini 4 models. Teams at Google and universities used it with human experts on open problems in mathematics, chemistry, software and machine learning. Google makes it available to other users.

How it works

The tool runs many tries at the same time. Checkers, such as the Lean 4 proof checker and software test suites, show which answers work. Each agent gets only the files and tools it needs.

The proof

Researchers from Google and three universities proved the Courtade and Kumar conjecture, from 2013, in full generality. Gemini 4 Argon made most of the new limits. An agent system also closed a problem on data analysis that was open since 2014. The authors checked the results.

What is not known

Google gives no cost data and no data on errors. Human experts framed the questions and edited the results for publication. Google says it will add more features in the next weeks.

Insight: Google checks agent work with tools, not trust

The tool pairs many tries with checkers that test each answer. Orchestrator, a related Google tool, makes two Gemini 4 models work in turns. Analyst Pareekh Jain says its documentation has no record of each step. This seems to show that Google puts checks first in research work, but not in coding work. Builders must use the same checks in code.