We study an experimentation problem where actions today have a long-lasting impact on information generation in the future. Actions are irreversible and generate information gradually over time. We solve for the optimal path of actions when the decision maker does not know the payoff-relevant state of the world. Because current choices have persistent effects, the problem has two state variables: a summary of past actions and the current belief on the state of the world. There is a novel informational trade-off as acting today speeds up information generation but postponing actions results in more informed choices. Our two leading examples cover the monopoly pricing of durable goods with social learning and capacity expansion in a market with uncertain optimal size. We show that since the monopolist can internalize future benefits from learning, the monopolist's optimal solution may result in a higher social surplus than the competitive market in both examples.

Most decisions– from a job seeker appraising a job offer to a policymaker assessing a novel social program– involve the consideration of numerous attributes of an object of interest. This paper studies the optimal evaluation of a complex project of uncertain quality by sampling a limited number of its attributes. The project is described by a unit mass of correlated attributes, of which only one is observed initially. Optimal sampling and adoption is characterized for both single-agent and principal-agent evaluation. In the former, sampling is guided by the initial attribute but it is unaffected by its realization. Sequential and simultaneous sampling are equivalent. The optimal sample balances the variability of sampled attributes with their usefulness in inferring neighboring attributes. Under principalagent evaluation, the realization of the initial attribute informs sampling so as to better influence adoption. Sampling hinges on (i) its informativeness for the principal, and (ii) the variation of the agent’s posterior belief explained by the principal’s posterior belief. Optimal sampling is not necessarily a compromise between the players’ ideal samples. I identify conditions under which mild disagreement leads to excessively risky or conservative sampling. Yet, drastic disagreement always induces compromise.