It cannot really oversee this. If you can decompose a problem into individual steps that are not, in themselves, against the agent's alignment, it's certainly possible to have the aggregate do so.