Abstract
While generative AI (GenAI) integration in education often remains at the tool-assistance level, this study proposes and evaluates an AI agent-based collaborative teaching system designed to operationalize structured human-AI interaction in vocational training. Utilizing a multi-agent architecture featuring role-specialized agents—customer simulation, decision support, and instructional guidance—the system implements a triadic framework that enables coordinated engagement among students, instructors, and AI agents across the entire learning cycle. A mixed-method longitudinal evaluation of a "Cross-Border E-Commerce Customer Management" course (2022–2025, N=232) revealed a significant upward trend in instructional quality, with average teaching evaluation scores rising from 90.58 to 93.81. These results demonstrate that agent-oriented design and iterative optimization effectively transform GenAI from an auxiliary tool into a collaborative educational partner. Despite current technical limitations in emotional intelligence, the proposed architecture offers a replicable, human-centered framework for interactive competency development and instructional optimization in vocational education and related skill-based domains.