This dissertation consists of three essays examining the agency issues in political economy. In Chapter 1, I analyze how a careerist delegate carries out reform decisions and implementation under alternative information environments. Regardless of his true policy preference, the delegate seeks retention and tries to signal to a principal that he shares an aligned policy predisposition. Given this pandering incentive, the principal best motivates the delegate's implementation if she can commit to a retention rule that is pivotal on reform outcomes. I characterize an ``informativeness condition'' under which this retention rule is endogenous, provided that the principal uses an opaque information policy -- she observes the delegate's policy choice and outcomes, but not the effort. With other information policies, the principal has to reward congruent policy choices rather than good policy outcomes; her policy interest is damaged by failing to sufficiently motivate reform implementation. In Chapter 2, I study the limits of mediated conflict resolution when: states have incentives to misrepresent private information about resolve; mediators have limited capacity to enforce agreement; and political leaders bargain in the shadow of audience costs and political bias. With a mechanism design approach, I characterize the conditions under which a mediator may propose a peaceful settlement to resolve the crisis. I find that the availability of peaceful settlements has more to do with political bias than audience costs. The reason is that, regardless of audience costs, the war payoff implied by political bias and the war technology imposes a lower bound that a particular state would ask for from any peaceful settlement. Absent peaceful settlements, I show that some stronger leaders would fight strictly more often and obtain strictly higher payoffs than others. Finally, I examine how the enforcement capacity may impact mediators' ability to implement peaceful settlements. In Chapter 3, I study a model of testing in which 1) a principal sets a private test standard to infer the type of an agent, and 2) the agent exerts unobserved influence on his/her test result. I characterize conditions under which the principal employs a threshold known as the ``reasonable doubt'' to trade off inference errors. Through comparative static analysis, the model offers insights into factors that impact the principal's equilibrium threshold. I find that different test natures, dichotomized by whether an innocent agent outperforms a guilty one when neither exerts influence, may guide comparative static predictions in the opposite directions. The analysis highlights the importance of probing the context and strategic concerns of testing, and generates empirical implications beyond the conventional wisdom that a decision-maker unambiguously leans towards the lesser of two harms.