Our analysis of misunderstandings is performed in a plan based architecture for generating and recognizing agent behavior, based on a two-level representation of the knowledge about acting. At the metalevel, the Agent Modeling (AM) library Ardissono-etal:96a describes the recipes for planning and executing actions; at the object level, there are the Domain and the Speech Act plan libraries, and some internal actions like those describing the interpretation task; the Domain library Ardissono-etal:93a describes the recipes for obtaining the domain goals in a restricted domain; the Speech Act Library Ardissono-etal:95a,Ardissono-etal:95d keeps the linguistic knowledge. The three plan libraries are based on a Generalization and a Decomposition Hierarchies Kautz:91 and share the same representation formalism, so that the same procedures can be used on them. Moreover, plans support a declarative representation style, that can be used both for interpreting and generating the behavior of the agent. The idea of using the same structures to model interpretation as well as generation is basic to our notion of coherence, that presupposes that agents have the same Agent Modeling plans (they have the same rational behavior, although their beliefs and goals may differ substantially). The agent's knowledge also includes some prior information about the world state; such knowledge is strictly domain dependent and regards beliefs about other agents, about conditions which hold in the world, etc. Although the presence of this type of knowledge poses a limit in the applicability of the model to general domains, it is a basic component of any system which aims at performing a deep interpretation of the interactions among agents.