This paper describes the PZ system, an assistant system that generates a personalized newspaper digest. To cope with the manifold user interests and the dynamic nature of the newspaper domain, the system is built as a multi-agent system. Each agent models a different facet of the user’s interest. Working together in a economy where only useful agents survive, the agents learn from the feedback the users provide as they read the digest. The use of implicit feedback was found to be sufficient for generating a good digest. Experiments with simulated users and experience with real users are reported.