Boris Cherny, the creator of Claude Code, tasked 12 AI agents with converting 64 scientific papers into interactive browser-based laboratories. The project aims to allow users to replicate experiments, adjust simulation conditions, and observe changes in real time, while viewing the underlying equations and code.
The agent team is divided into four specialized groups: readers extract equations and settings, builders convert methods into code, reviewers verify results against original data, and interface agents design the controls. Reviewers specifically check for system stability, mass conservation, and alignment with original measurements.
The project utilized 11 distinct methods and 28 reference scenes. According to the results, 7 experiments matched the reference data within an acceptable margin of error, while 4 experiments remain under review for potential discrepancies.
