In a network of information agents, the problem of how these agents keep accurate models of each other becomes critical. Due to the dynamic nature of information and the autonomy of the agents, the models that an agent has of its information sources may not reflect their actual contents. In this paper, we propose an approach to automatically reconcile agent models. First, we show how an agent can revise its models to account for the disparities with its information sources. Second, we show how to learn concise descriptions of the new classes of information that arise. Third, we show how the refined models improve both the accuracy of the knowledge of an agent and the efficiency of its query processing. A prototype for model reconciliation has been implemented using a SIMS mediator that accesses relational databases.