A fundamental and seemingly unbridgeable psychological boundary divides humans and nonhumans. Essentialism theories suggest that mixing these categories violates “natural kinds.” Perceptual theories propose that such mixing creates incompatible cues. Most theories suggest that mixed agents, with both human and nonhuman features, obligatorily elicit discomfort. In contrast, we demonstrate top-down, cognitive control of these effects—such that the discomfort with mixed agents is partially driven by disfluent categorization of ambiguous features that are pertinent to the agent. Three experiments tested this idea. Participants classified 3 different agents (humans, androids, and robots) either on the human-likeness or control dimension and then evaluated them. Classifying on the human-likeness dimensions made the mixed agent (android) more disfluent, and in turn, more disliked. Disfluency also mediated the negative affective reaction. Critically, devaluation only resulted from disfluency on human-likeness—and not from an equally disfluent color dimension. We argue that negative consequences on evaluations of mixed agents arise from integral disfluency (on features that are relevant to the judgment at-hand, like ambiguous human-likeness). In contrast, no negative effects stem from incidental disfluency (on features that do not bear on the current judgment, like ambiguous color backgrounds). Overall, these findings support a top-down account of why, when, and how mixed agents elicit conflict and discomfort.