This paper presents a new architecture of a classifier system for learning in virtual environments. The model will be integrated in our multi-user platform to provide interaction between intelligent agents and user clones. An agent is an autonomous entity equipped with sensors and effectors. Its behavior is guided by rewards coming from the environment that produce rules called classifiers. The knowledge is shared between agents by using the “sending-message” protocol to increase the global efficiency of the group. The classifier system is specially adapted to a multi-task environment and incorporates a short-term memory to record the recent events of the simulation. These ideas have been implemented and used to develop a virtual soccer where the user plays with autonomous agents that combine communication and evolution.