In this paper we present a computational approach to key aspects of understanding social interactions. First, we specify a class of problems – understanding fables – that require inference about agents’ mental states from their behavior. After this, we review earlier work on UMBRA, an abductive
system for single-agent plan understanding, and describe extensions that let it deal with multi-agent scenarios, including ones that involve accidental errors and intentional deceptions. These augmentations include distinguishing domain-level knowledge from more general content about social interactions and applying this knowledge at nested levels of belief. We also report the results of experimental studies on a set of fable-based scenarios that demonstrate the benefits of these extensions. In closing, we discuss how our approach to social cognition is informed by earlier research in the area.