This study employs the case study method. It traces how a junior high geography teacher uses a self-developed AI agent to design an interdisciplinary practical activity themed “Exploring Classroom West-Facing Sun Exposure.” Drawing on human-AI dialogue logs, iterative versions of the instructional plan, and teacher refl ection journals, this study reveals the decision-making mechanisms underlying human-AI collaborative design and summarizes a three-domain decision-making model. The findings reveal three distinct response behaviors exhibited by the teacher toward the AI-generated proposals: adoption, modifi cation, and rejection. Full adoption occurred exclusively in procedural tasks, thereby delineating the “effi cacy boundary” of the AI agent. Partial modifi cation focused on dimensions such as content supplementation, task hierarchy, and phrasing refi nement, refl ecting the core intervention point of teachers’ professional judgment. Total rejection was based on the presupposition of learning situation and resources, which revealed the cognitive blind spot of the agent. The study shows that human-AI design should not be a “Machine Output, Teacher Execution” model. Instead, teachers achieve professional leadership through revision. The value of an AI agent is not to replace teachers, but to stimulate their professional judgment. This study off ers micro-level evidence for understanding the mechanisms of human-AI collaborative design. It also inspires practical directions for both AI literacy training for teachers and the ongoing refi nement of educational agents.