Fungus-growing ants (Attini) are part of a complex symbiosis with Basidiomycetous fungi, which the ants cultivate for food, Ascomycetous fungal pathogens (Escovopsis), which parasitize cultivars, and Actinobacteria, which produce antibiotic compounds that suppress pathogen growth. Earlier studies that have characterized the association between attine ants and their bacterial symbionts have employed broad phylogenetic approaches, with conclusions ranging from a diffuse coevolved mutualism to no specificity being reported. However, the geographic mosaic theory of coevolution proposes that coevolved interactions likely occur at a level above local populations but within species. Moreover, the scale of population subdivision is likely to impact coevolutionary dynamics. Here, we describe the population structure of bacteria associated with the attine Apterostigma dentigerum across Central America using multilocus sequence typing (MLST) of six housekeeping genes. The majority (90%) of bacteria that were isolated grouped into a single clade within the genus Pseudonocardia. In contrast to studies that have suggested that Pseudonocardia dispersal is high and therefore unconstrained by ant associations, we found highly structured ([Formula: see text]) and dispersal-limited (i.e., significant isolation by distance; [Formula: see text], [Formula: see text]) populations over even a relatively small scale (e.g., within the Panama Canal Zone). Estimates of recombination versus mutation were uncharacteristically low compared with estimates for free-living Actinobacteria (e.g., [Formula: see text] in La Selva, Costa Rica), which suggests that recombination is constrained by association with ant hosts. Furthermore, Pseudonocardia population structure was correlated with that of Escovopsis species ([Formula: see text], [Formula: see text]), supporting the bacteria's role in disease suppression. Overall, the population dynamics of symbiotic Pseudonocardia are more consistent with a specialized mutualistic association than with recently proposed models of low specificity and frequent horizontal acquisition.