This paper develops a framework to assess how fear of miscoordination affects the sustainability of cooperation. Building on theoretical insights from Carlsson and van Damme (1993), it explores the effect of small amounts of private information on a class of dynamic cooperation games with exit. Lack of common knowledge leads players to second guess each other's behavior and makes coordination difficult. This restricts the range of equilibria and highlights the role of miscoordination payoffs in determining whether cooperation is sustainable or not. The paper characterizes the range of perfect Bayesian equilibria as the players' information becomes arbitrarily precise. Unlike in one-shot two-by-two games, the global games information structure does not yield equilibrium uniqueness.

This paper studies how agents with conflicting interests learn to cooperate when the details of cooperation are not common knowledge. It considers a repeated game in which one player has incomplete information about when and how her partner can provide benefits. Initially, monitoring is imperfect and cooperation requires inefficient punishment. As the players' common history grows, the uninformed player can learn to monitor her partner's actions, which allows players to establish more efficient cooperative routines. Because revealing information is costly, it may be optimal not to reveal all the existing information, and efficient equilibria can be path-dependent.