We use experiments to test comparative statics predictions of canonical tournament theory. Both the roles of principal and agent are populated by human subjects, allowing us to test predictions for both incentive responses and optimal tournament design. Consistent with theory, we observed an incentive effect from raising the winner's prize. However, we also observed several empirical puzzles that appeared to contradict theory. Controlling for social preferences did not resolve the puzzles, although social preferences do influence behavior. It turns out that the puzzles can be explained by the canonical model once the textbook assumption of separable agent utility is relaxed.