In this paper, we present a technique for developing user simulators which are able to interact and evaluate conversational agents. Our technique is based on a statistical model that is automatically learned from a dialog corpus. This model is used by the user simulator to provide the next answer taking into account the complete history of the interaction. The main objective of our proposal is not only to evaluate the conversational agent, but also to improve this agent by employing the simulated dialogs to learn a better dialog model. We have applied this technique to design and evaluate a conversational agent which provides academic information in a multi-agent system. The results of the evaluation show that the conversational agent reduces the time needed to fulfill to complete the the dialogs, thereby allowing the conversational agent to tackle new situations and generate new coherent answers for the situations already present in an initial model.