Speech-based training agents can be described as virtual humans posing as interactive training characters with the capability to communicate in a spoken conversational manner. While creating this technology, developers face two stumbling blocks: 1) modeling the agent and its training knowledge is a time-consuming and tedious task, and 2) modern speech recognition software suffers from high Word-Error Rates caused by noisy environmental conditions. This paper presents a dialog management architecture that addresses these problems using the Context-Based Reasoning paradigm. The system minimizes the time necessary to build the training knowledge in the instructional agent, as well as tolerates the relatively high Word-Error Rates related to automatic speech recognition. Ultimately, these advantages lead to quick development of speech-based training agents. The dialog manager was directly implemented into the LifeLike Avatar, an embodied conversational agent funded by the National Science Foundation. A set of quantitative results is presented to reflect the effectiveness of the system.