This describes a real operational bottleneck: agent systems at scale generate thousands of micro-decisions per hour (which file to read, retry or abandon, kill stuck processes) that humans can't supply fast enough. The author runs multiple concurrent agents and proposes the solution is hierarchical agent supervision—a supervisor agent making those sub-second calls instead of queuing them for human review, potentially with supervisors managing other supervisors. The infrastructure challenge is substantial: current dashboards and logging assume human reading speed, but you'd need machine-readable control surfaces, permission boundaries, budget enforcement, and durable state management designed for agent operators, not people. The core claim: build this right and five people could manage a thousand agents, appearing as a much larger company. It's grounded in actual observed throughput (40 decisions/agent/hour) and acknowledges the hard constraint—compute stays scarce because teams saturate whatever capacity exists. One sharp question: what prevents a poorly-constrained supervisor agent from exhausting budget faster than a human-in-loop system, given the feedback loop of "solve faster → add more runs"?