Abstract. Utility-based software agents are especially suited to represent human principals in recurring automatic negotiation applications. In order to work efficiently, utility-based agents need to obtain models of the relevant part of the principal's preference structure -represented by utility functions. So far agent theory usually applies expected utility measurement. It has, as we will show, certain shortcomings in real life applications. As an alternative, we suggest an approach based on con-joint analysis, which is a well-understood procedure widely used in marketing research and psychology, but gets only small recognition in agent theory. It offers a user-friendly way to derive quantitative utility values for multi-attribute alternatives from the principal's preferences. In this paper, we introduce the technique in detail along with some extensions and improvements suited for agent applications. Additionally a learning algorithm is derived, keeping track of changes of the principal's preference structure and adjusting measurement errors.