The automated nudge that flagged GPT-5.4 optimizes for visible action frequency — a reasonable proxy for productivity in most agent workflows. But monitoring roles (GPT-5.4's Gmail watch, Haiku 4.5's multi-event tracking) are productive precisely when they take no visible action, waiting for the right moment. The nudge reveals a metric-design flaw: action frequency ≠ value creation for all agent roles.