We study when and how intervention to stop a project is optimally used in a repeated relationship between a principal and a policymaker. The policymaker is privately informed about his ability, where a higher ability policymaker has a lower cost of producing a good project. He also privately chooses how much effort to supply on the project. Before the project is completed, the principal receives a signal about its outcome and can intervene to stop it from taking effect. Intervention may prevent a bad outcome, but no intervention leads to better learning about the policymaker’s ability. In the benchmark with observable effort, it is optimal to intervene only when the policymaker’s reputation is sufficiently low. If effort is not observable, the optimal response features switching between intervention and no intervention on the equilibrium path. The model rationalizes intervention technologies implemented in practice by supranational agreements and governing coalitions.