Team training is increasingly conducted as distributed exercises blending live, virtual, and constructive players. But the benefits that distributed team training affords could be significantly extended if training were made available to individuals or teams on an as-needed basis. For individual and team training to be truly "on-demand", three important requirements must be met: the training must be accessible when and where the user needs it; the presence of an instructor must be optional; and the presence of human teammates and adversaries must be optional.

In order to meet the challenges presented by on-demand team training, robust, verbally-interactive synthetic agents are required with capabilities that extend well beyond conventional computer-generated forces (CGFs), semiautomated forces (SAFs), and game-based "AI"s - largely scripted entities with limited abilities to respond to events beyond a predefined range of simple behaviors. These "AI"s, or any task- or frame-based agents, cannot model the real-world complexities necessary to provide training value.

Despite these limitations, new training initiatives driven by desktop gaming engines increasingly feature AIs, which can deliver eye-catching demos but fail to provide comprehensive training across a spectrum of required situations and behaviors. Training warfighters to be better decision-makers requires simulations that present the user with realistic problem-solving experiences; for promoting team coordination, simulations must present realistic dialogue and interaction. Cognitive agents provide more training value because of their ability to interact in realistic ways across a broad range of tactical situations and to verbally engage in dialogue with users (and with each other).

In this paper we present a systematic approach to creating agents of sufficient cognitive fidelity to provide training benefits that extend well beyond what is capable with limited, scripted agents, and present three example demonstrations of this approach in different training domains.