High-level deliberation in robotic systems has to handle three different but closely interwoven aspects simultaneously: intended tasks, contingencies, and system-level errors. To reduce the complexity, we propose the system modes concept to abstract runtime state information and reconfiguration of the software components of the underlying layers by a model-based approach. The proposed concept introduces a notion of hierarchically composed, virtual subsystems as well as a notion of modes that determine their configuration. It features an inference engine to deduce the modes of the subsystems from the components and top-down reconfiguration mechanisms. Repetitive and fine-grained communication between high-level deliberation and underlying software components can thereby be reduced, decreasing unwanted coupling between system parts.