A multi-agent retrieval-augmented generation framework for pedagogical feedback on EFL student writing

High-quality feedback can support writing development, yet producing it consistently at classroom scale remains difficult. Large language models generate fluent comments quickly, but their feedback can over-emphasise surface grammar, drift from the target rubric, assert errors not present in the essay, or disclose complete rewrites. We propose RUMAR, a retrieval-augmented framework for English-as-a-Foreign-Language writing feedback. RUMAR combines error detection and rubric-aware retrieval with a bounded Generation–Evaluation–Refinement loop comprising two language-model agents—a Generator and a Critic—and a deterministic refinement controller. The Critic evaluates candidate feedback against five explicitly anchored 1–5 criteria, and an aggregator organises accepted comments across Hattie’s feed-up, feed-back, and feed-forward layers. On the reported ASAP-AES and ELLIPSE evaluation, RUMAR has the highest observed point estimate in each automatic-metric column relative to the six implemented baselines. Descriptive expert ratings and a four-week trial with 80 B1–B2 learners at one institution also favour RUMAR among the automated systems compared. These findings are limited to the reported datasets, argumentative tasks, and automated comparators. The trial contains neither a human-teacher nor a no-feedback control, and the one-week delayed post-test cannot establish durable transfer. We therefore interpret the results as evidence of feasibility and pedagogical alignment within the evaluated setting, not as proof of superiority over teacher feedback or broad classroom readiness.