Human intelligence is being increasingly redefined to include the all-encompassing effect of emotions upon what used to be considered 'pure reason'. With the recent progress of research in computer vision, speech/prosody recognition, and bio-feedback, real-time recognition of affect could very well prove to enhance human-computer interaction considerably, as well as to assist further progress in the development of new emotion theories. We propose an adaptive system architecture designed to integrate the output of various multimodal subsystems. Based upon the perceived user’s state, the agent can adapt its interface by responding most appropriately to the current needs of its user, and provide intelligent multi-modal feedback to the user. We concentrate on one aspect of the implementation of such an environment: facial expression recognition. We give preliminary results about our approach which uses a neural network.