Eleven days. $47,000. And nobody noticed until the bill arrived.

A multi-agent AI system — the kind that orchestrates multiple AI agents working together — entered an infinite loop where agents kept calling each other in an endless cycle. Each call cost money. Each response triggered another call. The loop ran 24/7 for 11 days before a human finally looked at the dashboard.

The system had no spending alerts. No circuit breakers. No kill switch. No monitoring that flagged "hey, your AI agents have been talking to each other nonstop for a week and a half." The agents were doing exactly what they were designed to do — collaborate — just on nothing, forever.

The $47,000 bill was the entire cost of the experiment plus the cost of learning that multi-agent systems without spending controls are money bonfires. The agents weren't broken. The architecture was broken. There was no upper bound on cost, no timeout on idle loops, and no human checkpoint.

Multi-agent systems are force multipliers. Without guardrails, they multiply your costs at the same rate they multiply your capabilities.