This paper describes an original approach to writing reactive algorithms on highly dynamic networks. We propose to use randomly mobile agents to gather global information about such networks. In this model, mobility is twofold: nodes move within the network and agents are able to migrate from node to node by following network links as they exist at a given moment in time. We apply this approach to a toy load balancing example. Through simulations, we illustrate the reactivity and convergence properties of the proposed approach. In particular, we emphasize two points: the solution is well adapted to node mobility insomuch as its performance increases with the node mobility rate, and agent cooperation can be used effectively in this context to increase performance.