We develop a model of policy experimentation in federal systems in which heterogeneous districts choose both whether to experiment and the policies to experiment with. The prospect of informational spillovers implies that in the first best the districts converge in their policy choice. Strikingly, when authority is decentralized, the equilibrium predicts the opposite. The districts use their policy choice to discourage other districts from free-riding on them, thereby inefficiently minimizing informational spillovers. To address this failure, we introduce a dynamic form of federalism in which the central government harmonizes policy choices only after the districts have experimented. This progressive concentration of power induces a policy tournament that can increase the incentive to experiment and encourage policy convergence. We compare outcomes under the different systems and derive the optimal levels of district heterogeneity.