Video-based classroom analysis offers great opportunities for teachers to reflect on their own or others’ performance for professional development. However, the substantial time and resource demands of expert or peer insights restrict teachers’ access to consistent instructional support. Large language models (LLMs) have the potential to generate content-specific comments based on the classroom transcript, but their insights often lack novelty and expert professional vision. To address this, we propose an LLM-based comment generation mechanism through multi-agent reflection, which simulates expert instruction by integrating professional vision and rubric-based feedback. Experimental results demonstrate that this mechanism can effectively generate insightful comments with deeper analyses of student thinking and more specific suggestions for teaching improvement, facilitated through iterative revisions between LLM agents in the reflection process. The study presents an effective method for generating insightful feedback for teachers by simulating expert guidance, which is expected to enhance the practical value of generative AI for teacher coaching.