Does granting agents more autonomy undermine human oversight?

"AI Agents Push Humans Out of the Loop" is a position paper against treating a "human in the loop" as a simple fix for agent risk. Its claim is that "current approaches to the development and deployment of AI agent systems do not support effective human oversight – they contribute to its degradation." The introduction makes the opening move: governance frameworks, vendors and policymakers have converged on human oversight, "however, the presence of an overseer does not entail reliable oversight." The discussion states the conclusion bluntly: "In the current state of the art, oversight degrades the overseer."

The paper names two channels. The first is position: "The more autonomy agents are granted, the less the user is positioned to oversee them." The second is capacity: the cognitive capacities oversight requires, which the paper lists as situational awareness, critical judgment and domain skill, "are undermined by the act of using these systems." The two compound. Users end up approving plans they have not meaningfully reviewed, accepting rationales they have not independently evaluated and certifying actions whose consequences they cannot anticipate, until the human is "a superficial actor in a system they cannot meaningfully penetrate." The remedy the abstract proposes is to treat the overseer's human needs "at the same level of importance as AI agent capability," through design-level affordances and organizational protocols that support critical judgment and counteract skill atrophy from extended use of automation.

This adds a cause to a limit the vault already holds. Can organizations lose scrutiny capacity while keeping oversight forms? says the loop can stay on paper after the capacity to scrutinize is gone, but it states this as something organizations "can" do and names no mechanism. Here the erosion is attributed to the agent systems themselves: the tool the overseer uses is the thing wearing the overseer down, and the trend is described as continuing as agents are deployed "at greater scale and across more consequential domains." It also qualifies Should AI systems stay collaborative rather than fully autonomous?, whose prescription of human involvement assumes the human keeps the ability to correct and clarify. And it reads the finding in Does AI risk increase with the autonomy we give it? from the oversight side: the paper does not argue that more autonomy widens what one failure can touch, but that it weakens the check meant to catch the failure. Does targeted human oversight beat both full autonomy and exhaustive review? asks where to put the human. This paper asks whether the human placed there can still judge, which is a separate question.

What the excerpt does not give. It is the abstract plus one introduction passage and one discussion passage, so no study, sample, measure or rate of skill atrophy appears; the supporting literature is cited only by bare reference numbers. It does not say which affordances or protocols it proposes, or how well any of them work. The degradation claim is therefore best held as a position argued from the automation and human-computer interaction literature, not a result measured on current agent deployments. What follows at that strength is a design obligation: a system that leans on a human reviewer should say what keeps that reviewer's situational awareness, judgment and skill intact, instead of inferring oversight from the reviewer's presence.