Abstract: To address the problems of insufficient autonomous inquiry, ambiguous human-machine responsibility boundaries, inadequate review of AI-generated results, and weakened collaboration that may occur in educational agent-assisted experimental teaching, this paper identifies four types of risks across the stages of experiment preparation, experiment implementation, collaborative inquiry, and outcome evaluation, namely responsibility shifting, review absence, collaboration dilution, and passive design. It also designs and implements a four-dimensional closed-loop process supervision approach throughout pre-class, in-class, post-class, and after-class activities, transforming students’ operational responsibilities and review obligations in human-machine collaboration into observable and evaluable teaching behaviors. Taking programming experiment teaching in Object-Oriented Curriculum Design as a typical case, this study conducted practical verification. Multi-source process data from 76 students in the autumn semester of 2025 show that the AI-generated code annotation rate and cross-validation execution rate increased by 35% and 28%, respectively; the accuracy rate of ethical scenario tests reached 93%; and students’ ability to distinguish between tool suggestions and personal decisions was improved. The results indicate that clarifying the auxiliary role of educational agents, embedding process review procedures, and strengthening evidence recording in human-machine collaboration help maintain students’ autonomous inquiry and sense of responsibility, providing a reference for risk prevention and quality evaluation in AI-assisted experimental teaching.