This paper describes an approach to user adaptation realized in a multiagent interface system for interaction with a virtual environment. The interface agency adapts to users’ individual preferences by learning from direct feedback. The core idea is that agents that were succesful in meeting the user’s expectations are given credit while unsuccesful agents are “discredited.” Communicating credit values, agents organize themselves so that the overall behavior of the interface agency gradually adapts to the individual user as the session is proceeding.