Abstract—Asynchronous discussion forums are a key pedagogical feature of MOOCs and online degree programs, yet assessing the quality of student engagement at scale remains a persistent challenge. This paper explores the use of large language models (LLMs) to automatically detect cognitive presence and summarize cognitive engagement in online discussions. Grounded in the Community of Inquiry framework, we examine how LLMs can support instructors in understanding the depth of student thinking across large enrollments. We analyzed over 1,500 discussion posts from an introductory MOOC and a graduate-level online computer science course. Using LLaMA-based models, we implemented classification pipelines to categorize posts by cognitive presence phases (triggering event, exploration, integration, and resolution). We compared single-agent and multi-agent architectures where LLMs cross-validate or critique each other’s output to evaluate model accuracy, reliability, and interpretability. Our best-performing models achieved over 92% agreement with human coders. In addition to classification, we employed LLMs to summarize the cognitive trajectory of entire discussion threads, surfacing patterns in student reasoning and critical thinking. Findings suggest that LLMs can provide scalable, accurate insights into students’ cognitive engagement, with both single- and multi-agent systems showing promise for enhancing model robustness. We discuss the pedagogical implications of integrating these tools into digital learning environments and outline key considerations for model generalizability. This study contributes to the growing field of AI-enhanced learning analytics, offering a practical framework for automated feedback in large-scale online education.