Building general-purpose image segmentation systems is a central concern for many researchers. The challenge is of course to propose more and more flexible and reliable systems, the behaviour of which closely depends on the task to perform and the images to process. In this paper, a knowledge-based system for the dynamic building of image segmentation applications is described. This system is based on the monitoring of a library of image segmentation operators. First, the distinctive features of the segmentation domain are studied, in order to find out which problem solving paradigm is the most relevant in this case - hierarchical planning - and seven behavioral rules for an architecture dedicated to solving segmentation problems are proposed. Our implementation is based on the Blackboard model. Cooperation and knowledge acquisition issues, which are essential in our approach, are also briefly discussed.