Recent advancements in Large Language Models (LLMs) have demonstrated their potential in automating pedagogical content generation. However, the design and evaluation of high-quality lesson plans still require iterative refinement and pedagogical alignment. In this paper, we propose AgentLesson, a multi-agent framework that simulates human-like collaboration between a Writer Agent and an Evaluator Agent to iteratively generate and improve lesson plans. Given a subject, grade, and topic, the Writer Agent creates an initial plan based on Gagné’s Nine Events of Instruction, while the Evaluator Agent provides rubric-based feedback across five dimensions: Clarity, Integrity, Depth, Practicality, and Pertinence. This feedback is used to guide the Writer Agent in multiple rounds of refinement. We conduct extensive experiments across multiple subjects and topics, and quantitatively evaluate plan quality under single-agent, one-round, and multi-round agent settings. Our results show that multi-agent collaboration significantly enhances lesson quality, with notable improvements in Depth and Pertinence. This work highlights the promise of LLM-based agent collaboration in educational content creation and opens new directions for teacher-assistive technologies.