Agricultural extension programs often train a few farmers and count on diffusion through social networks for the innovation to spread. However, if markets are imperfectly integrated, this may also inflict negative externalities. In a two-step experiment of an agronomy training program among Rwandan coffee farmers, we first randomize the concentration of trainees at the village level and then randomly select within each village. Knowledge increased, and yields were 6.7% higher for trained farmers. We find no evidence of social diffusion; instead, control households experienced negative spillovers in high treatment concentration areas, likely because of competition for a scarce input, fertilizer.