Models of agenda setting are central to the analysis of political institutions. Elaborations of the classical agenda‐setting model of Romer–Rosenthal have long been used to make predictions about policy outcomes and the distribution of influence among political actors. Although the canonical model is based on complete and perfect information about preferences and policy outcomes, some extensions relax these assumptions to include uncertainty about preferences and reversion points. We consider a different type of uncertainty: incomplete knowledge of the mapping between policies and outcomes. In characterizing the optimal agenda setting under this form of uncertainty, we show that it amends substantively the implications of the Romer–Rosenthal model. We then extend the model dynamically and show that rich dynamics emerge under policy uncertainty. Over a longer horizon, we find that agenda control suppresses the incentive of legislators to experiment with policy, leading to less policy learning and worse outcomes than are socially efficient.