Sources
OpenAI's Thibault on Ultrafast Mode: Incident Commanders Get It During Outages, and It Cuts Parallel Agents From Fifteen to Four
In a 45-minute interview posted 2026-08-24, Matthew Berman's guest from OpenAI describes internal ultrafast-mode allocation going to the incident commander and response team during outages, with the rest reserved for customers. Berman describes his current workflow as kicking off 10 to 15 parallel agents and waiting 30 to 45 minutes per task, and both agree fast inference collapses that to three or four agents you actually supervise. The guest also names skill files as the clunkiest part of coding harnesses today, saying people have found them hard to maintain over time and that memory is still thin.
↳ Follow the thread