Most current virtual world systems focus on the interaction between a single agent and the user. This simplification does not reflect the richness of a real social environment. The quantitative increment from the simple two-party interaction to a multi-party interaction does not merely increase the difficulty linearly. In fact, it leads to a much more complex situation involving multimodal communication, utterance understanding, and interaction style. Here we introduce a four-layer agent architecture with multiparty interaction support. A Newtonian law learning environment based on this agent architecture is presented and how multiple agents cooperate to improve user learning is illustrated. The agent's interaction ability within a multiparty environment can be realized in three sections: planning and task execution, communication and understanding, as well as learning and coaching. Our entire system can be regarded as a step toward addressing and solving issues related to effective teaching in a multi-user environment within a sophisticated domain.