Due to their promise for supporting applications society cares about and their unique blend of distributed systems and networking issues, wireless sensor networks (SN) have become an active research area. Most current SN use an arrangement of nodes with limited capabilities. Given SN device technology trends, we believe future SN nodes will have the computational capability of today's handhelds, and communication capabilities well beyond today's 'motes'. Applications will demand these increased capabilities in SN for performing computations in-network on higher bit-rate streaming data. We focus on interesting fusion applications such as automated surveillance. These applications combine one or more input streams via synthesis, or fusion, operations in a hierarchical fashion to produce high-level inference output streams. For SN to successfully support fusion applications, they will need to be constructed to achieve application throughput and latency requirements while minimizing energy usage to increase application lifetime. This thesis investigates novel middleware mechanisms for improving application lifetime while achieving required latency and throughput, in the context of a variety of SN topologies and scales, models of potential fusion applications, and device radio and CPU capabilities. We present a novel architecture, DFuse, for supporting data fusion applications in SN. Using a DFuse implementation and a novel simulator, MSSN, of the DFuse middleware, we investigate several middleware mechanisms for managing energy in SN. We demonstrate reasonable overhead for our prototype DFuse implementation on a small iPAQ SN. We propose and evaluate extensively an elegant distributed, local role-assignment heuristic that dynamically adapts the mapping of a fusion application to the SN, guided by a cost function. Using several studies with DFuse and MSSN, we show that this heuristic scales well and enables significant lifetime extension. We propose and evaluate with MSSN a predictive CPU scaling mechanism for dynamically optimizing energy usage by processors performing fusion. The scaling heuristic seeks to make the ratio of processing time to communication time for each synthesis operation conform to an input parameter. We show how tuning this parameter trades latency degradation for improved lifetime. These investigations demonstrate MSSN's utility for exposing tradeoffs fundamental to successful SN construction.