To address the challenge of maintaining learning continuity while providing personalized instruction in large-scale industrial software education, this paper proposes a collaborative multi-agent pedagogical framework. The framework establishes a four-layer agent collaboration architecture—comprising instructional coordination, skill decomposition, context management, and evaluation feedback—and introduces a “three-layer, multi-attribute, and relational” skill ontology model to enable dynamic granularity decomposition of complex engineering tasks. By integrating a hierarchical memory system and a tool invocation framework, the system achieves deep coupling with industrial software APIs and establishes a real-time, closed-loop verification mechanism spanning from operational execution to automated diagnosis. Results from a quasi-experimental study demonstrate that the proposed framework outperforms traditional instructional modes in terms of learning gains, instructional efficiency, and knowledge retention rates, while effectively providing differentiated scaffolding support based on the learners’ initial proficiency levels. This research provides a systematic solution for industrial software talent cultivation, ranging from skill modeling to intelligent intervention, and offers significant practical reference value for the development of intelligent pedagogical platforms integrated with industry-education synergy.

Conference: 15th International Conference on Educational and Information Technology (ICEIT 2026) Date: March 27-29, 2026 Location: Xi’an, China