It is commonly recognized that a key challenge of urban military operations is the complex terrain. However, the large number and diversity of people present another less recognized but critical challenge, especially for noncombat operations. Warfighters must understand the socio-cultural effects of their actions, handle direct interaction with non-combatants (and potential combatants), and understand the immediate and long-term consequences of their interactions: every soldier must be a skilled ambassador. Given that realistic training requires many participants, the high cost and limited availability of highly-qualified participants restricts access to quality training opportunities. Synthetic characters can provide the numbers, but present-day simulated characters have limited capabilities for interacting with each other and with humans in the rich, meaningful way that is required. Thus, we must populate these virtual training worlds with characters whose behaviors are generated by human behavior models (HBMs) rather than humans. While significant achievements have been made in developing HBMs that are able to control a single simulated character (or a single group of simulated characters), a serious limiting factor has been the inability of heterogeneous HBMs to interact with each other or share knowledge about their environment. We present an architecture and multi-level message framework for enabling HBMs to communicate with each other about their actions and their intentions. Also, we describe our conception of the "Mindscape," which will facilitate the use of massive numbers of synthetic characters by representing the options and state of social relationships and interactions as shared affordances in the environment. Finally, we describe an application of our prototype training system in which three heterogeneous HBMs developed by various organizations interoperate within a single training-oriented scenario. We believe that this approach will encourage the development of standards for interoperability among HBMs that will lead to the development of richer training and analysis simulations.