We compare evaluations of employee performance by individuals and groups of supervisors, analyzing a formal model and running a laboratory experiment. The model predicts that multi-rater evaluations are more precise than single-rater evaluations if groups rationally aggregate their signals about employee performance. Our controlled laboratory experiment confirms this prediction and finds evidence that this can indeed be attributed to accurate information processing in the group. Moreover, when employee compensation depends on evaluations, multi-rater evaluations tend to be associated with higher performance.