AI agents run in multi-step loops, demanding a fundamentally different infra stack

sarahookr · x · 2026-09-02

Sarah Guo argues agents are reshaping inference workloads: historical AI workloads are predictable continuous matrix multiplication, while agents operate in multi-step loops—execute an action, check for success, adjust course. Like a cookie recipe requiring repeated oven in-and-out, the bottleneck shifts from getting dough to the oven to the check-and-reinsert cycle. This requires a fundamentally different infrastructure setup.

Related event: Sarah Hooker: Test-Time Compute and Agents Are Forcing an AI Infrastructure Rebuild(9 posts)→

Original post →

More from coding & agent

coding & agent channel →