Google and MIT introduce Coco agent platform for hardware and ML model co-design

Researchers from Google and MIT have introduced Coco, an agent platform built to support hardware and model co-design for TPU architects, detailed in a new [research paper](https://arxiv.org/abs/2610.02376). The system addresses workflow environments where design decisions rely on hundreds of gigabytes of simulation sweeps not present during model pretraining.

Coco records simulation sweeps in a normalized relational database so reported figures derive directly from SQL queries. Its agents can autonomously compose typed tools, perform recurring analyses such as comparing systems at matched execution configurations, and use the navigation state of the user interface as context. The near-term objective is to halve the time architects spend preparing simulations and extracting insights.