The intent of incentive-based contracts -- which tie compensation to performance in professions like teaching -- is to improve productivity. In practice, the effects of such contracts have diverged markedly from predictions. The intent of this dissertation is to expand contract theory and provide empirical evidence from both the laboratory and real-world incentive programs on how contracts in complex environments, such as teaching, may be substantially improved. An innovation of this work is to present a theoretical model that considers the effects of output-based incentives when agents lack knowledge of the production function. In the context of incentive contracts for teachers, I expand on contract theory by adding uncertainty around the marginal productivity of inputs -- such as different classroom activities -- towards student test outcomes. I test the theoretical predictions of this model using variation in the implementation of evaluations in the Washington DC teacher incentive program and in the setting of a laboratory experiment. In my first paper, "Personnel Contracts under Production Uncertainty: Theory and Evidence from Teacher Performance Incentives," I test the prediction that, due to production uncertainty, teacher incentives based on in-class evaluations may be substantially more effective than test-based incentives by separately identifying how two types of teacher incentives affect student outcomes. In the IMPACT program, teachers can be fired or receive large bonuses based on a combination of observational measures in unannounced in-class evaluations -- which can be thought of as measures of teacher inputs -- and test-based measures of the effect of teachers on student outcomes. I measure how teachers modify their behavior when they have no threat of an evaluation, and how those changes affect student test scores. Because the timing of in-class observations is random, the assignment of treatment -- how many days a teacher has the threat of an evaluation -- is exogenous. I find that increasing the number of days without the possibility of an evaluation leads to a decline in students' tested scores, which is inconsistent with a model in which agents know the production function, but consistent with my model of production uncertainty. I demonstrate that all of the positive effects of the IMPACT program can be explained entirely by the effect of a possible in-class evaluation, suggesting the test-based incentive has little or no effect. A takeaway from this analysis is that incentive-based compensation targeting production inputs may yield significant gains in the effectiveness of incentive contracts. In my second paper, "Teacher Improvements in Windows of High-stakes Observation," I look deeper at the specific behavioral changes caused by the IMPACT incentive in Washington DC. I map how teachers modify specific components of their practice, as measured on their evaluations, in response to the daily probability of an in-class evaluation. In so doing, I illustrate the predictions of the Holmstrom-Milgrom multi-tasking model by showing teachers make the most improvements on teaching practices that are easily adjusted first. In the standard multi-tasking model, if employees devote all their attention on a single incentive, overall productivity may fall. However, overall teacher responses to the possibility of an evaluation still induce meaningful improvements in student outcomes, demonstrating the small cost of Holmstrom-Milgrom-style multi-tasking relative to the large gains from reducing employee production uncertainty by using an input-based incentive. In my third paper, "Teacher Performance Pay through the Lens of Production Uncertainty: Theory and Evidence from a Real-Effort Laboratory Experiment," I test theoretical predictions of the production uncertainty model in a laboratory setting, which allows for controlled randomization in the production function in order to causally identify the effects of production uncertainty. I imitate uncertainty in the marginal value of inputs -- analogous to inputs for student test scores -- by asking participants to solve easy or hard problems to earn financial rewards, but the marginal payoffs are drawn from known distributions. Treatments vary by changing the variance of the marginal payoff to each task. I find that, as predicted, increased production uncertainty induces participants to favor inputs with lower variance in marginal productivity, even while holding all other things constant (including average marginal payoff).