Coco AI platform deployed with TPU architects for hardware and model design

Researchers report deploying Coco, an AI agent platform, with TPU architects to help set up experiments, run simulators and analyze designs for machine learning models and the accelerators that execute them. The platform registers simulation results in a relational database so agents retrieve numbers through SQL queries. Typed tool interfaces let agents combine operations, while specialized workflows compare systems at matched execution configurations, including designs outside the best-performing tradeoff frontier. The preprint describes early deployment experience aimed at shortening simulation and analysis work, without quantifying a time reduction.