The idea of coevolution originated with Darwin's proposal that long-proboscid pollinators and long-tubed flowers might be engaged in reciprocal selection, but this has not been demonstrated. Here we test key aspects of Darwin's hypothesis of reciprocal selection in an experiment with naturally interacting populations of extremely long-proboscid flies (Moegistorhynchus longirostris: Nemestinidae) and long-tubed irises (Lapeirousia anceps: Iridaceae). We show that the benefit derived by both the fly (volume of nectar consumed) and the plant (number pollen grains received) depends on the relative length of their interacting organs. Each trait is shown to act both as agent and target in directional reciprocal selection, potentially leading to a race. This understanding of how fitness in both species varies in relation to the balance of their armament allows us to make tentative predictions about the nature of selection across multiple communities. We find that in each community a core group of long-tubed plant species might together be involved in diffuse coevolution with the fly. In poorly matched populations, the imbalance in armament is too great to allow reciprocal selection to act, and these species might instead experience one-sided selection that leads to convergence with the core species. Reciprocal selection drives the evolution of the community, then, additional species become attached to the network of interacting mutualists by convergence.