Behavioral acts constitute the building blocks of interpersonal perception and the basis for inferences about personality traits. How reliably can observers code the acts individuals perform in a specific situation? How valid are retrospective self-reports of these acts? Participants interacted in a group-discussion task and then reported their act frequencies, which were later coded by observers from videotapes. For each act, observer-observer agreement, self-observer agreement, and self-enhancement bias were examined. Findings show that (a) agreement varied greatly across acts; (b) much of this variation was predictable from properties of the acts (observability, base rate, desirability, Big Five domain); (c) on average, self-reports were positively distorted; and (d) this was particularly true for narcissistic individuals. Discussion focuses on implications for research on acts, traits, social perception, and the act frequency approach.