We study the intertemporal properties of accounting conservatism with a focus on managerial incentives. In our main model, conservatism results in smaller expected payouts to the manager (agent) in early periods and larger expected payouts in later periods. Conservatism shifts (ambiguous) evidence that might be used to recognize good performance in early periods to later periods. In later periods, good performance is less informative, since good news might mean good current period performance and might also mean good prior period performance whose recognition was delayed. Because of the intertemporal shift in the information content of performance measures, incentives provided in future periods can spillback to early periods, making conservatism preferred by the principal (shareholders). We also study an extension in which the principal learns about the firm over time. In the learning model, conservatism is optimal because it reduces the expected payment in the first period when providing incentive is more difficult than it is in the second period. In a final extension, we study overlapping projects that give rise to a multi-task setting. In the multi-task setting, unbiased accounting is preferred to a maximally biased one (conservative or aggressive) when bias introduces so much measurement heterogeneity that one of the otherwise identical projects becomes a significant "bottleneck." When the projects are productively heterogeneous and one of the projects would be a bottleneck without bias, conservative bias can make the two projects more similar from a measurement perspective, eliminating the bottleneck.