We study a dynamic principal-agent setting in which both sides learn about the importance of effort. The quality of the agent’s output is not observed directly. Instead, the principal jointly designs an evaluation technology and a wage schedule. More precise performance evaluation reduces current agency costs but promotes learning, which is shown to increase future agency costs. As a result, the optimal evaluation technology is both imprecise and tough: a bad performance is always sanctioned, but a good one is not always recognized.

We also study the case in which principal and agent have different priors, for instance because the agent is overconfident. Then, the principal uses a tough evaluation structure to preserve the agent’s profitable misperception. For an underconfident agent, by contrast, she either uses a fully informative evaluation in order to promote learning and eliminate costly underconfidence, or is lenient if learning is too costly.

We investigate the impact of transparency (and other information policies) in settings where, consistent with empirical evidence, agents underinfer from equilibrium statistics. We adapt cursed equilibrium to a beauty contest game. To discipline information acquisition with mislearning, we propose a subjective envelope condition which allows for a tractable analysis while maintaining behaviorally plausible assumptions: agents correctly anticipate their actions but incorrectly deem them optimal. We show that this condition characterizes the rest points of a simple learning process.

Cursed agents use and acquire more private information, creating a positive externality. Welfare increases for low degrees of cursedness, as these gains exceed the losses from incorrect use. Transparency crowds out private information but always increases welfare. Policies targeting fundamental information may backfire as they distract cursed agents from a source of information they already underuse. We also investigate the behavior and welfare of an atomistic rational agent in a cursed economy.

We study a decision-framing design problem: a principal faces an agent with frame-dependent preferences and designs an extensive form with a frame at each stage. This allows the principal to circumvent incentive compatibility constraints by inducing dynamically inconsistent choices of the sophisticated agent. We show that a vector of contracts can be implemented if and only if it can be implemented using a canonical extensive form, which has a simple high-low-high structure using only three stages and the two highest frames, and employs unchosen decoy contracts to deter deviations.

We then turn to the study of optimal contracts in the context of the classic monopolistic screening problem and establish the existence of a canonical optimal mechanism, even though our implementability result does not directly apply. In the presence of naive types, the principal can perfectly screen by cognitive type and extract full surplus from naifs.

We propose a tractable framework to introduce externalities in a screening model. Agents differ in both payoff-type and influence (how strongly their actions affect others). Applications range from pricing network goods to regulating industries that create externalities. Inefficiencies arise only if the payoff-type is unobservable. When both dimensions are unobserved, the optimal allocation satisfies lexicographic monotonicity: increasing along the payoff-type to satisfy incentive compatibility, but tilted towards influential agents to produce the externality. In particular, the allocation depends on a private characteristic that is payoff-irrelevant for the agent. We characterize the solution through a two-step ironing procedure that addresses the nonmonotonicity in virtual values arising from the countervailing impact of payoff-types and influence. If observable, influence is used as a signal of the payoff-type. We provide sufficient conditions for rents from influence to emerge even in a setting featuring atomistic agents.