Summary form only given. Commodity clusters have enabled ambitious multiphysics or coupled modeling of complex, mutually interacting, computationally intensive systems in science and engineering. Each individual sub-system is represented as a component with its own parallel processor layout and requirements for temporal advance. A central challenge in developing such systems is the parallel coupling problem, which involves overall system architecture and the automation of component registration, distribution of the processor pool between individual components, parallel data transfer and transformation. There currently exist efficient mechanisms for automating parallel data transfer and transformation such as MCT and MPCCI. Mechanisms for top-level system integration, including component registration and resource allocation, scheduling, and control at runtime are less mature and face even greater challenges in heterogeneous environments. We will discuss the numerous architectural choices faced in framework and parallel coupled application development, and will illustrate them through a comparison of these mechanisms in four scientific application frameworks: the community climate system model, the space weather modeling framework, the earth system modeling framework, and the weather research and forecasting model. We will then discuss a more sophisticated set of requirements for automating these functions in application frameworks for heterogeneous clusters and computational grids