To create the most effective possible simulations, domain experts must be able to author, monitor, and modify the behavior of simulated agents. Current computational models of autonomous agent behavior are not adequate in this regard. Simple hard-coded models still predominate in many areas, while the most capable and realistic behavior modeling architectures - such as SOAR and ACT-R - are also generally the most difficult to work with, requiring trained programmers to develop and update behavior models. We contend that to enable domain experts without programming expertise to author sophisticated agent behaviors, there are two main challenges that must be addressed: condition authoring and behavior analysis.

Complex conditions - such as the preconditions for a step in a plan - are a necessary part of almost any behavior model, but specifying these conditions is not easy. Text-based authoring is an efficient way to enter the information, but the required syntax can be overwhelming to the non-programmer. Visual authoring methods, by contrast, are better able to guide non-programmers through the authoring process but tend to be much more time-consuming and laborious. The second major challenge is enabling non-programmers to analyze the runtime behavior of the models they create. Behavior models of any significant complexity require multiple "test and fix" iterations to uncover authoring mistakes. Modeling tools must therefore provide data visualizations that permit the non-programmer to see both global structure and specific details in the large volume of data generated by test runs of the behavior model. In addition, authoring tools must easily allow the creation of unit-test-like scenarios.

We have spent the last three years developing an adversary behavior modeling tool for the Air Force, during which time we have attempted to address both of these challenges. We will present lessons learned and suggested best practices as well as areas for future work.