This work presents an agent-based interface that is not merely reactive to some user request, but is proactive since it is capable of engaging in a goal-directed conversation with the user, e.g., by taking the initiative to recommend new products. The naturalness of interaction, especially for casual users, is enhanced by appropriate 2D facial models. The proactiveness of the agent is based on a recommendation engine that combines content-based retrieval, which exploits user profiles based on content features extracted from the dialogue and descriptions of items that users find relevant, with collaborative filtering, which clusters users according to their expressed taste to generate recommendations within these virtual communities. The proposed system has been evaluated and validated by using a top-down approach, focusing on the system/user interaction.