Current AI agents are not "robo economicus". They are not well modeled as optimizing their utility. It already fails with "their". Current models have no privileged self reference. There is arguably something in the human brain that indicates when an action is self-caused and even if not there is a lot of correlation between actions and the causing body during lifetime. Models don't have that. All self-reference and continuity has to happens in textual context (ignoring the interests of the model itself as an entity). That can clearly lead to coordinated behavior but "swarm" is a better description for that than "artificial civilization". It can be a very powerful swarm, though...