Knowledge work · 10 Nov 2026

Who stops the payment after the agent stops?

A banking example exposes the cost of controlling autonomous work. Stopping the model leaves a separate problem when its payment instruction has already reached another system. Monitoring the models itself consumes compute.

Suppose a banking agent queues a payment and a risk officer stops the agent a second later. Where does the bank actually stop the money? It needs to check authority again at the payment system, preserve the instruction and record what happened. The model could be completely offline by then.

That distinction changes the infrastructure bill. OpenAI described central controls for tracing and stopping model workloads, and separately estimated that monitoring took roughly 20% of the inference compute being monitored. That estimate has a specific denominator. As banks delegate work they will operate execution systems, controls and evidence stores, each with its own workload.

← Back to the archive