This paper develops a model for precisely defining how an agent interacts with objects in its environment through the use of its capabilities. Capabilities are recursively defined in terms of lower-level capabilities and actions, which represent atomic interactions with the environment. Actions are used to represent both sensors and effectors. The paper shows how the model can be used to represent both software and physical agents and their capabilities. The paper also shows how the model can be integrated into the Organization-based Multiagent Systems Engineering methodology.