The Resource That Is Not on the List: Human Attention in an Operating System Layer for Agentic AI

An operating system layer for agentic AI is being specified right now, and its resource lists are converging: compute, memory, network, cost, latency, tokens, accelerators. The resource we argue bounds it is not on that list. One researcher now runs agents that generate at laboratory scale. Generation has scaled with the models; judging has not. Judging is a person's and not the layer's to enlarge, so what a layer changes is how many decisions each validated result costs, and that number is absent in two connected ways: absent as a managed resource, and absent from the record. Three measurements support this, over 54 papers read for content, 46 of them coded against items we fixed before searching, each a case where an AI system generated candidates and at least one was validated, and the reference titles of 24 more. The first asks what these proposals treat as a resource: across the six papers nearest that question, no paper both measures a person's load and throttles what produces it, since those that measure act only on where an item goes and those that throttle name no person. The second asks what applications record about the step where a person selects: of the 46, five report enough for a success rate to be computed at all, and the three counts that describe the funnel are reported by 16, 15 and 11. We claim the level and not the ordering, which at this sample size is not separable from noise. Of fourteen full texts, one attaches a number to a time saving. The third asks whether the field cites the literature on review capacity: across the 1,305 recovered reference titles, the vocabulary of enforcement and of human factors appears zero times, against 23 for a comparably specific methods vocabulary from an adjacent field that this corpus does cite. The cost is not unmeasurable: adjacent fields measure it, including a coefficient for how acceptance falls as decisions per occasion rise. We close with five requirements for counting that cost and lowering it without hiding it. The three we read have one of the five, part of a second, and none of the joints between them. We write the five as an interface contract with a checker and take four of them to a case the layer must refuse, written before the layer could pass it. All four failed first, in our own code, written by the same people who specified them.