We report the results of a field experiment with bicycle messengers in Switzerland and the United States. Messenger work is individualized enough that firms can choose to condition pay on it, but significant externalities in messenger behavior nonetheless give their on-the-job interactions the character of a social dilemma. Firms therefore suffer efficiency losses when messengers fail to cooperate. Second-mover behavior in our sequential Prisoner's Dilemma allows us to characterize the cooperativeness of our participants. We find that messengers, like our student controls, have heterogeneous social preferences, but are much more cooperative than students. Among messengers, we find that employees at firms that pay for performance are significantly less cooperative than those who are paid hourly or are members of cooperatives. To examine whether the difference is the result of treatment or selection we exploit the fact that firm type is location-specific in Switzerland and that entering messengers must work in performance pay firms in the U.S. We find that the erosion of cooperation under performance pay is predominantly due to treatment, and that the treatment effect is relatively rapid, more akin to the differential cueing of a behavioral norm than the gradual acquisition of a new preference.