Abstract We investigate David Lewis’ signalling games dynamics in the context of neural agents as embedded in realistic graph structures. We follow in the footstep of both the deep learning tradition—in moving away from pure rule-based models—, and the symbolic literature—in leveraging discrete, top-down structures to constrain the learning process. Through a series of experiments in which we systematically vary the social graphs connecting the players, we are able to show for the first time that the dynamics of language emergence within a population of neural agents is strongly influenced by the underlying graph topology constraining their interactions.