Model-driven applications may maintain large networks of structured data models and transformations among them. The development of such applications is complicated by the need to reflect on the whole network any runtime update performed on models or transformation logic. If not carefully designed, the execution of such updates may be computationally expensive. In this paper we propose a reactive paradigm for programming model transformations, and we implement a reactive model-transformation engine. We argue that this paradigm facilitates the development of autonomous model-driven systems that react to update and request events from the host application by identifying and performing only the needed computation. We implement such approach by providing a reactive engine for the ATL transformation language. We evaluate the usage scenarios that this paradigm supports and we experimentally measure its ability to reduce computation time in transformation-based applications.