Wireless sensor network (WSN) applications are often characterised by close coupling between their software and hardware components, which may result in ad-hoc, platform-specific software, together with the loss of portability and evolvability. We introduce a fractal composition-based approach for constructing and dynamically reconfiguring WSN applications. The approach uses π-calculus semantics to unify the models of interaction for both software and hardware components, on both local and remote nodes. Applications are constructed by forming compositions of interacting components, and can be decomposed and reconfigured into different topologies. The advantages of the approach are that it reduces the complexity of WSN programming; results in portable and evolvable software; and allows changes to be managed during execution without having to take the system off-line. We present an outline of the approach, and illustrate it with an example specified in the Insense language.