Smart agents represent users by creating models of particular users, and the agents' recommendations are based on these models. But models are rarely complete; this problem is particularly acute when there are changes either in the tastes and preferences of the individual, or in the marketplace (e.g., new products). Our agent attempts to solve this by actively trying to learn a user's preferences. The active agent balances two goals: to be immediately useful and to make high-quality recommendations ("selling"); and to learn more about the user. These recommendations have a lower quality, but also a higher "up-side" potential ("learning").