From one side we want to approve agent's intentions before it starts acting. From another side, we don't want to micromanage it.