When encountering difficulties in conventional educational games, learners seldom self-regulate to discover and organize the learning content in the game environment. With the development of the human-computer interaction technology, computer agents are gradually being applied to educational games to provide personalized guidance or support to learners in specific situations. Different roles of computer agents may affect students' learning achievements and feelings in the learning environment. Therefore, to compare the effects of different roles of computer agents on students' learning outcomes in digital games, this study developed a Tutor Agent-based Digital Game (TA-DG), a Peer Agent-based Digital Game (PA-DG) and a Student Agent-based Digital Game (SA-DG). The experimental results showed that the TA-DG and PA-DG learners had better learning achievements than the SA-DG learners. The PA-DG and SA-DG learners had higher motivation than the TA-DG learners. The PA-DG learners had a higher critical thinking tendency than the other two groups. In addition, the SA-DG learners had the lowest cognitive load, followed by the TA-DG and PA-DG learners. This implies that the individual roles of the agents had different impacts during the digital game-based learning process. The findings provide a good reference for developing future digital game-based learning research with personalized support.