Quantum Neural Monitoring of Agentic Swarms: Separating the Monitor from the Workers

We present a design for overseeing agentic artificial-intelligence systems: programs that call tools, open sockets, and can leave a sandbox faster than a person can read a log. Most organizations now plan to deploy such agents. Few can oversee them at that speed. A human dashboard is not fast enough to be the first control. A second copy of the same classical model is not a monitor either. That copy can be manipulated through its prompts, can share files or sockets with the workers, or can be retrained on the incident it was asked to stop. At that point the supposed monitor is only another worker. We specify the monitor as a quantum neural network (QNN) whose internal Hilbert space is divided into a protected core, a control plane, and an interface wall. The QNN is required to finish one cycle inside one worker step. The QNN may read telemetry from the workers and from the wall. The QNN must not copy those readings into its core parameters. We call that forbidden update write-back. A human remains in the loop as a session, with authority to override and to walk away, and is notified after a halt. The human is not the isolation clock. The wall instrument issues the halt. The halt freezes network exits, file mounts, sockets, credentials, and accelerators, and leaves the machine running for audit. This is a design paper. The manuscript reports no fleet incident and no new hardware measurement.