A multi-agent system is described where each agent consists of two kinds of neural networks. The agents interact by obtaining certain actions of the other agents as their inputs and generating own actions as their output. The actions of the different agents are mutually adjusted. The result of these adjusting processes is a social order that defines a certain group structure. This order is an emergent result from the agents interactions; in this sense the system evolves its order by itself. One may call this procedure “emergent programming”. When new agents are added to the group they will be “typified”, i.e., the newcomers are getting the same status as that agent that is most similar to them. Such emergent programming methods can be applied to internet agents or robotics where agents have the task to adjust to situations that could not be planned beforehand.