Software agents can prove useful in representing the interests of human users of agent systems. When users have diverse interests, the question arises as to how agents representing their interests can be grouped so as to facilitate interaction between users with compatible interests. This paper describes experiments in the DIET (Decentralised Information Ecosystem Technologies) agent platform that use evolutionary computation to evolve preferences of agents in choosing environments so as to interact with other agents representing users with similar interests. These experiments suggest a useful way for agents to acquire preferences for formation of groups for information interaction between users, and may also indicate means for supporting load balancing in distributed systems.