Everyone agrees agents need "human oversight," but in practice it ranges from rubber-stamping to full manual review that defeats the purpose of automation. For agents taking real actions — spending money, sending messages, modifying systems — what oversight design actually catches mistakes without becoming a bottleneck?