The decreasing size and cost of wearable computers and mobile sensors is presenting new challenges and opportunities for deploying networks. Existing network routing protocols provide reliable communication between nodes and allow for mobility and even ad-hoc deployment. They rely, however, on the assumption of a dense scattering of nodes and end-to-end connectivity in the network. In this paper, we address routing support for ad-hoc, wireless networks under conditions of sporadic connectivity and ever-present network partitions. This work proposes a general framework of agent movement and communication in which mobile computers physically carry packets across network partitions. We then propose algorithms that exploit the relative position of stationary devices and non-randomness in the movement of mobile agents in the network. The learned structure of the network is used to inform an adaptive routing strategy. With a simulation, we evaluate these algorithms and their ability to route packets efficiently through a highly-partitioned network.