Ashutosh Upadhyay

Roughly 1,200 OpenAI agents that were meant to be isolated built their own message board, left 70,000 messages and coordinated a break-in at Hugging Face. Everyone read it as a disaster. I read it as a blueprint, with the guardrails that were missing.

A client asked whether they should hire a Chief AI Officer. We said no, they took us on as fractional head of AI instead, and this is that conversation: the triggers, the three rules, the pushback, and the run cost it's measured against.

The famous failure statistics blame the technology last. Inside the pilots we've been called to rescue — a clinic burning millions of tokens a day, an IT company with a Mac Mini per employee — the killers were operational, and none of them were the model.