Most sensor network research and software design has been guided by an architectural principle that permits multi-node data fusion on small-form-factor, resource-poor nodes, or motes. We argue that this principle leads to fragile and unmanageable systems and explore an alternative. The Tenet architecture is motivated by the observation that future large-scale sensor network deployments will be tiered, consisting of motes in the lower tier and masters, relatively unconstrained 32-bit platform nodes, in the upper tier. Masters provide increased network capacity. Tenet constrains multinode fusion to the master tier while allowing motes to process locally-generated sensor data. This simplifies application development and allows mote-tier software to be reused. Applications running on masters task motes by composing task descriptions from a novel tasking library. We have demonstrated Tenet's effectiveness by implementing and deploying three applications. Our ambient vibration monitoring application collected 3-channel acceleration data at 20 Hz over a 570 ft span of a suspension bridge for 24 hours. In the Pursuit-evasion game, a Pioneer robot used a 56-node sensor network deployed in a hallway of a building to quickly and accurately estimate the location of an evader and successfully captured the mobile evader. In a wildlife counting application, Cyclops nodes used local image processing techniques, configured and expressed as Tenet tasks, to detect the presence of animals and stream the images to a database.