A new AI tool called Coco is helping TPU chip architects sift through mountains of simulation data without a chatbot that guesses when it does not actually know the answer.
Coco is an agentic platform built around four layers, according to a paper describing its deployment with TPU architects. A datastore pipes every simulation sweep into a normalized SQL schema instead of leaving results scattered across heterogeneous files. A library of typed-API tools lets agents compose their own analysis steps without a human stitching things together by hand. Purpose-built agents encode recurring workflows, including "iso-execution analysis," which compares chip designs at matched execution configurations, even ones that got swept but landed off the Pareto frontier. A fourth layer, the platform's UX, tracks an architect's navigation state and feeds that back to the agents as context, so the tool knows what you were just looking at.
The plumbing exists to dodge a specific failure mode. Hardware-software co-design decisions rest on hundreds of gigabytes of fresh simulation results that no model was trained on and that appear nowhere in outside literature. A generic chat-with-your-data bot would have to hallucinate its way through exactly the numbers that need to be correct. Coco's fix is to force every claim through a SQL query against real runs, which is a duller and more trustworthy approach than letting an LLM reason freely over data it has never seen.
The team reports early gains in time-to-simulation and time-to-insight, but those are impressions from a live deployment, not published benchmarks. Until actual numbers show up, treat this as solid infrastructure, not proof that chip design just got faster.