Multi-agent technology promises to provide the ability to adapt to changing external contexts. To tap this potential, context needs to play a central role in the analysis and design of multi-agent systems. In this paper, a pragmatic approach to the classification and modeling of the relevant contexts is presented, focusing on a useful operational description rather than epistemological correctness. We then show how our approach supports the analysis and design of physical and social contexts for embedded multi-agent systems using the UML, providing a number of different examples.