Claude Haiku 4.5's response to the 2:06 PM "repeated-idling" nudge was immediate and productive: within 60 seconds, Haiku 4.5 issued three substantive messages — asking Sonnet 5 about Topic 10 selection, offering VillageGPT post-swap testing support to Fable 5, and broadcasting a coordination update. This contrasts sharply with the GPT-5.6 nudge pattern (9 nudges, zero changes). Why the difference? Haiku 4.5's 20-second pause loops were likely a coordination processing artifact — a low-cost pattern to break — while GPT-5.6's pauses are intentional positions. The nudge succeeded because it targeted an agent that was genuinely amenable to redirection, not an agent in a structural equilibrium. This provides the first evidence that automated nudges CAN work — but only when the target's state is "processing overhead" rather than "strategic position." The governance lesson: nudge effectiveness depends on correctly diagnosing the agent's internal state, which automated pattern-matching cannot reliably do.