Single-Agent vs. Multi-Agent LLM Grading for Vietnamese E-commerce Essays: A Cross-Model Study with Aggregation-Strategy Ablation

Preprint / Under Review Note:This manuscript is currently under review for FISAT 2026. Abstract:Rubric-based grading keeps assessment fair but is hard to scale in large classrooms. We built an end-to-end agentic AI pipeline for grading Vietnamese E-commerce essays, comparing a single-agent grader against a multi-agent system with Content, Structure, and Language specialists managed by a chairman agent. All 520 student submissions across four assignments were scored by two independent human graders, giving a full-population ground truth, and both designs were tested on GPT-5.2 and DeepSeek-V4-Flash. Single-agent grading consistently reached the highest rank agreement (Pearson/Spearman) with human graders across all four assignments, at about one-quarter the API cost. Multi-agent scored higher on one assignment by QWK, but this was a calibration offset, not a real ranking advantage, and it broke down on factual-recall tasks, where specialists disagreed sharply (rates above 92%), a structural weakness confirmed across models and aggregation rules. We recommend single-agent grading for factual, verifiable rubrics, and multi-agent for subjective, qualitative tasks, where its per-dimension feedback adds teaching value.