AI agents are gaining ever greater autonomy, which introduces substantial risks. A widely suggested mitigation is to keep a “human in the loop,” meaning that humans supervise the agents, but this is far from straightforward: current agent designs often hinder effective oversight, and prolonged reliance on automation can erode the very cognitive capacities required for supervision. This position paper argues that present approaches to developing and deploying AI agents neither support effective human oversight nor prevent its degradation. Consequently, advancing AI agents must prioritize supporting the situated goals and cognitive needs of overseers, treating these human requirements as equally important as improving agent capabilities. To operationalize this, the paper links automation and human‑computer interaction research to AI agent workflows, outlining design‑level affordances and organizational protocols that (1) enable overseers to exercise critical judgment and (2) counteract skill atrophy caused by extended automation use. Developers and deployers are urged to adopt these or similar strategies. Without explicit support for the cognitive demands of effective human‑agent interaction, AI agent systems will continue to passively undermine the human skills they depend on.
Blogger's Review: The paper serves as a timely reminder that technological progress must not come at the expense of human supervisory abilities. Embedding human cognitive needs into the design of AI agents is essential for building safe and reliable systems.