The goal of the Soar/IFOR project is to provide intelligent agents capable of replacing human agents in large-scale distributed military simulations and smallscale, focused training exercises. The need for computer generated agents to remain reactive determines the requirements of the NL capability: (1) it nmst occur ill real time, (2) it nmst seamlessly integrate with the agent’s non-linguistic capabilities, e.g. perception, planning, reasoning about the task, and (3) its content nmst be comprehended in accordance with performance data, i.e. with all of the idiosyncratic constructions, ungrammaticalities, and self-corrections found in real language. Within the context of these research issues, we introduced NL-Soar, a language comprehension and generation capability designed to provide integrated, real-time natural language processing for systems built within the Soar architecture [Lewis, 1993; Nelson el al., 1994a; Nelson et al., 1994b; Rubinoffand Lehman, 1994]. At present, our work has focused on addressing the issues in (1) and (2) and this paper report on the solutions found.