Attention-Constrained Oversight for Agentic AI: Risk, Reversibility, Authority, and Human Attention

As artificial intelligence systems move from recommending actions to executing them, human oversight can no longer be treated as a binary choice between autonomy and a human-in-the-loop. This paper proposes an attention-constrained framework for designing agent autonomy around four variables: risk, reversibility, authority, and human attention. Risk, reversibility, and authority describe the consequence profile of an action; human attention describes the scarce oversight resource available to review it. Drawing on research on levels of automation, appropriate reliance, automation bias, meaningful human oversight, and recent empirical studies of software agents, the paper argues that frequent approval requests can reduce rather than increase meaningful control when they consume attention faster than humans can exercise judgment. The framework introduces an action-level oversight demand, an explicit human-attention budget, a lightweight extension for cumulative exposure across action sequences, and two design principles: exception-based oversight and approval quality over approval count. It further distinguishes four complementary modes of oversight - boundary setting, monitoring, escalation, and audit - and shows how they can be combined so that routine, bounded actions proceed without synchronous review while consequential or difficult-to-reverse actions receive concentrated human attention. The paper also links these modes to task-level delegation contracts and proposes adaptive safeguards when review conditions appear to degrade. The contribution is integrative: reliable human oversight is framed not as maximizing human presence, but as allocating human judgment where its marginal value is highest.