We study the design of approval rules when experimentation must be delegated to an agent with misaligned preferences. Our motivating example looks at how the FDA can design approval rules a function of the outcome clinical trials. In these clinical trials, the agent (the drug company) must pay a cost for experimentation and may have information about the likelihood that the state is high (the drug is good).
We study this question in a dynamic learning framework and look at how the level of commitment the regulator can place on the agent changes the structure of the optimal approval rule, both when the agent has private information about the payoff relevant state and when he does not. When the mechanism must satisfy only ex-ante participation constraints, the optimal approval rule becomes a stationary threshold (similar to the problem with no agency concerns). However, when the mechanism must satisfy interim, participation constraints, the approval threshold will no longer be stationary change over time. We find the optimal approval rule and show that it moves downward monotonically. Surprisingly, the approval threshold only moves downward as a function of the minimum of the regulator's beliefs. When the agent possess private information about the state, we find that the agent with high information may receive a fast-track: his approval threshold is initially low (the fast-track) but takes a jump up if the regulator's beliefs fall too low. These dynamics are to our knowledge new and help us understand how approval rules change over time.