Can AI create genuinely deep things?Can I trust agent-written code without reading every line?Can good interface design be constrained and checked?