This paper proposes a multi-agent workflow, WikiHowAgent, that leverages large-scale language models (LLMs) to simulate interactive teaching-learning conversations. WikiHowAgent integrates teacher and learner agents, an interaction manager, and an evaluator to facilitate procedural learning and assess educational quality. We present a dataset of 114,296 teacher-learner conversations spanning 17 domains and 727 topics, based on 14,287 tutorials. We use an evaluation protocol that combines computational and criteria-based metrics with human judgment alignment. We demonstrate the effectiveness of the workflow in various settings and provide insights into the capabilities of LLMs across domains. The dataset and implementation are fully open-source.
