The effectiveness of digital game-based language learning (DGBLL) has been recognized by scholars. With the development of computer technology and multimedia learning environments, computer agents have been widely used in game systems to provide learning guidance or assistance. A computer agent is a virtual character created in a digital learning system to achieve instructional goals. However, in traditional teaching systems, computer agents are designed to perform a single role, such as a teacher or a student. Computer agents with a single interactive logic often lack the functionality required to understand students' conditions and needs from various perspectives, and cannot adapt for better learning support. Therefore, this study proposed an adaptive role-switching strategy that allows computer agents to adjust their roles and functions according to students' needs in digital games to promote their learning achievement. An adaptive computer agent-based digital game was developed to investigate the impact of this model on English vocabulary learning achievement, motivation, self-efficacy, and English anxiety among EFL (English as Foreign Language) students. A quasi-experiment was designed and 56 middle school students in four classes were recruited. Two classes (n = 30) were arranged to be the experimental group which used an adaptive computer agent-based digital game (adaptive computer agent-based digital game, ACA-DG), while two classes (n = 26) were arranged to be the control group which used the conventional computer agent-based digital game (conventional computer agent-based digital game, CCA-DG). The results showed that students in the experimental group had significantly higher learning achievement and self-efficacy than those in the control group. On the other hand, students in the experimental group had significantly lower English anxiety than those in the control group. There was no significant difference between the two groups regarding learning motivation.