That last one! If there’s any conceivable justification for turning an unpredictable AI agent loose with no human-gating on its actions, it’s the confidence that your permissions structure prevents it from doing any real damage. And then you just…hand it all the permissions?