How do political actors learn about their environment when the “data” provided by political processes are characterized by rare events and highly discontinuous variation? In such learning environments, what can theory predict about how learning actors will take costly actions that are difficult to reverse (e.g., eliminating programs, approving a risky new product, revising a security policy, firing or recalling an appointed or elected official)? We develop a formal model for this problem and apply it to the termination of bureaucratic agencies. The conventional wisdom that “the older a bureau is, the less likely it is to die” (Downs 1967, Inside Bureaucracy) persists but has never been properly tested. This paper offers a learning-based stochastic optimization model of agency termination that offers two counterintuitive predictions. First, politicians terminate agencies only after learning about them, so the hazard of agencies should be nonmonotonic, contradicting Downs's prediction. Second, if terminating agencies is costly, agencies are least likely to be terminated when politicians are fiscally constrained or when the deficit is high. We assess the model by developing a battery of tests for the shape of the hazard function and estimate these and other duration models using data on U.S. federal government agencies created between 1946 and 1997. Results show that the hazard rate of agency termination is strongly nonmonotonic and that agencies are less likely to be terminated under high deficits and divided government. For the first 50 years of the agency duration distribution, the modal termination hazard occurs at five years after agencies are enabled. Methodologically, our approach ties the functional form of a hazard model tightly to theory and presents an applied “agenda” for testing the shape of an empirical hazard function. With extensions, our model and empirical framework are applicable to a range of political phenomena.