To address the pain points in traditional software engineering teaching, such as disconnected learning scenarios and limited effectiveness of project-driven teaching, this study designs and implements a multi-role AI Agent named SE_Coach. It simulates roles in software development, such as product managers, and guides students through heuristic questioning and situational simulation. A 16-week controlled experiment was conducted with 80 junior computer science students. The experimental group adopted a model of teacher guidance + AI agent + group collaboration, while the control group used traditional group collaboration. The results show that the experimental group had significantly better self-efficacy and project outcome quality. SE_Coach effectively improves students’ practical abilities and collaborative thinking, providing a feasible paradigm for AI-empowered software engineering teaching.