Abstract: Organizations are riddled with cooperation problems, that is, instances in which workers need to voluntarily exert effort to achieve efficient collective outcomes. To sustain high levels of cooperation, the experimental literature demonstrates the centrality of reciprocal preferences but has also overlooked some of its negative consequences. In this paper, we ran lab-in-the-field experiments in the context of open-source software development teams to provide the first field evidence that highly reciprocating groups are not necessarily more successful in practice. Instead, the relationship between high reciprocity and performance can be more accurately described as U-shaped. Highly reciprocal teams are generally more likely to fail and only outperform other teams conditional on survival. We use the dynamic structure of our data on field contributions to demonstrate the underlying theoretical mechanism. Reciprocal preferences work as a catalyst at the team level: they reinforce the cooperative equilibrium in good times but also make it harder to recover from a negative signal (the project dies). Our results call into question the idea that strong reciprocity can shield organizations from cooperation breakdowns. Instead, cooperation needs to be dynamically managed through relational contracts.