Google & MIT's Coco: an agent platform for TPU hardware-model co-design
dair_ai · x · 2026-10-07
A Google–MIT report introduces Coco, an agent platform TPU architects use for hardware/model co-design. Every decision draws on hundreds of GB of fresh simulation sweeps absent from pretraining; all sweeps live in a normalized relational DB so agent-reported numbers come from SQL queries. Agents compose typed tools, run recurring matched-configuration analyses, and read UI navigation state as context. Near-term goal: halve the time architects spend setting up simulations and extracting insights.
More from coding & agent
- 55,000 Lines From One Prompt: Matt Shumer Open-Sources the Gauntlet Loop Method — letandrewcook · 2026-10-07
- Developer uses Opus 5.5 to recreate seven Adobe apps including Photoshop as free open-source Rust alternatives — kimmonismus · 2026-10-07
- Adding t_wada's commenting principles to AGENTS.md slashed noisy AI code comments — 4310sy · 2026-10-07
- openTPU: A One-Person Open-Source AI Accelerator Designed by AI Agents, Running 10 LLMs on an FPGA — ai · 2026-10-07
- Agent workflows don't matter — your log of traces for training future orchestrators is what compounds — natesiggard · 2026-10-07
- "Fix your failing self until the task completes": a developer's go-to Codex prompt — jdjohnson · 2026-10-07