Training negotiation is difficult because it is a complex, dynamic activity that involves multiple parties. It is often not clear how to create situations in which students can practice negotiation or how to measure students' progress. Some have begun to address these issues by creating artificial software agents with which students can train. These agents have the advantage that they can be “reset,” and played against multiple times. This allows students to learn from their mistakes and try different strategies. However, these agents are often based on normative theories of how negotiators should conduct themselves, not necessarily how people actually behave in negotiations. Here, we take a step toward addressing this gap by developing an agent grounded in a cognitive architecture, ACT-R. This agent contains a model of theory-of-mind, the ability of humans to reason about the mental states of others. It uses this model to try to infer the strategy of the opponent and respond accordingly. In a series of experiments, we show that this agent replicates some aspects of human performance, is plausible to human negotiators, and can lead to learning gains in a small-scale negotiation task.