This work explores the benefits of using user models for plan recognition problems in a real-world application. Self-interested agents are designed for the prediction of resource usage in the UNIX domain using a stochastic approach to automatically acquire regularities of user behavior. Both sequential information from the command sequence and relational information such as system's responses and arguments to the commands are considered to typify a user's behavior and intentions. Issues of ambiguity, distraction and interleaved execution of user behavior are examined and taken into account to improve the probability estimation in hidden Markov models.