Grids are evolving towards supporting a dynamic, collaborative, multidisciplenary problem-solving environment coined the virtual organization (VO). To date, adaptivity has been treated in an ad-hoc fashion leading to application-specific solutions that cannot be reused. In this proposal, a framework for exposing, exploiting, and coping with adaptivity in Grids is proposed. The key elements of this framework are the definition of OGSA-compliant adaptive Grid service classes for high-end services common to parallel/distributed scientific computing as well as adaptive resource provider services. To ease the implementation of adaptive Grid services, reusable middleware will be designed as part of this project. This next-generation middleware addresses four fundamental issues required to construct adaptive services: (1) adaptive service interfaces, (2) dynamic performance prediction, (3) adaptive resource management, and (4) adaptive run-time infrastructure for MPI. Several novel approaches are proposed including algorithms for dynamically selecting performance estimators as well as selecting resource management strategies.