From Debate to Writing: Insights from a Multi-Agent Debate System in Enhancing Argumentative Writing

Argumentative writing requires learners to engage in critical thinking through sustained reasoning, perspective-taking, and evidence-based judgment. Although generative artificial intelligence is increasingly used in writing instruction, limited attention has been given to how structured multi-agent interaction supports the reasoning processes underlying high-quality writing. This study examines a role-based multi-agent debate system embedded in an English for Academic Purposes course. A total of 47 undergraduate students participated in a sequential instructional design progressing from single-agent practice to debate-to-writing tasks within the multi-agent debate environment. Results show significant improvements in English proficiency, with stronger gains in writing performance. Interaction data indicate that students engage in sustained dialectical exchanges with AI debaters rather than relying on surface-level tutoring. Toulmin-based analysis of students’ written texts shows generally high levels of writing quality following engagement with the multi-agent debate system, particularly in sustaining claim-data structures, while the development of counterargument and rebuttal remains uneven. Interviews further suggest that the system supports students in engaging with multiple perspectives, alongside challenges in coherence and language use. Overall, the findings suggest that the effectiveness of the multi-agent debate system lies not simply in providing knowledge, but in structuring dialogic processes that support the development of argumentation.