A position paper from Margaret Mitchell and colleagues makes a structurally important argument: human oversight of AI agents is not a fixed floor that systems can rely on, but a resource that erodes over time as agents grow more capable, more autonomous, and more deeply embedded in workflows. The paper frames this degradation as a predictable dynamic, not an edge case — as agents handle more steps without interruption, the human's ability to meaningfully review, redirect, or recover erodes in proportion.
For UX designers, this reframes the design problem considerably. Oversight isn't something you bolt on as a compliance feature; it requires active, ongoing investment in interfaces that keep humans genuinely capable of intervening — not just nominally in the loop. The paper's argument implies that patterns like approval, interruption, and review will atrophy in practice unless they are deliberately designed to remain usable as agent capability scales.
The practical implication is that oversight UI has a kind of half-life. An approval gate that made sense when agents handled two-step tasks may become a rubber-stamp ritual once agents operate across dozens of actions. Designers need to think about how controls stay meaningful — and how to signal to users when their oversight has effectively become theater.