Abstract. Task-Oriented Computing supports the realization of user needs through the automatic composition of services from service descriptions and user tasks, i.e., high-level descriptions of the user needs. Service-Oriented Computing supports the description, publication, discovery and composition of services. Yet, automatic service composition processes commonly assume that service descriptions and user tasks share the same abstraction level, and that services have been pre-designed to integrate. To release these strong assumptions and to augment the possibilities of composition, we add adaptation features into the service composition process using semantic descriptions and adaptive extensions to graph planning. Our adaptive composition technique is fully automatic and has been implemented in a prototype tool.