In this paper, we describe our experience with the design and implementation of an embodied conversational agent (ECA) that converses with users in order to change their dietary behavior. Our intent is to develop a system that dynamically models the agent and the user and adapts the agent's counseling dialog accordingly. Towards this end, we discuss our efforts to automatically determine the user's dietary behavior stage of change and attitude towards the agent on the basis of unconstrained typed text dialog, first with another person and then with an ECA controlled by an experimenter in a Wizard of Oz study. We describe how the results of these studies have been incorporated into an algorithm that combines the results from simple parsing rules together with contextual features using a Bayesian network in order to determine user stage and attitude automatically.