Розробка багатоагентної генеративної структури конвеєрів для генерації навчальних планів з детермінованою перевіркою обмежень

Large language models (LLMs) are increasingly used to generate structured learning plans aligned with outcome-based education (OBE). The object of the study is a multi-agent workflow for generating a structured OBE learning-plan package with a final deterministic verification stage. The problem addressed is the low reliability of LLM-generated outputs, which frequently violate schema rules, numeric constraints, and cross-artifact consistency requirements. To solve this problem, a multi-agent generative pipeline is proposed, decomposing the task into six specialized agents followed by deterministic constraint verification applied to the final artifact bundle. Structural reliability is measured using completeness and compliance, while cross-artifact coherence is evaluated through redundancy, spacing, phase progression, and assessment fit. The evaluation involves 12 courses with 10 repeated runs per course (120 runs per variant) across four different LLMs to assess cross-model robustness. The results show that the multi-agent pipeline achieves completeness of 0.9682–1.00 and compliance of 0.9376–0.9698, significantly outperforming the single-agent configuration (completeness 0.5926–0.6580; compliance 0.4698–0.4853). These improvements are explained by task decomposition, which reduces structural failure propagation, and deterministic verification, which rejects invalid outputs and preserves referential integrity. Ablation analysis indicates that the Course Character agent exerts the highest impact on overall performance. The proposed framework can be applied in higher education curriculum planning under OBE conditions, using minimal course metadata and producing machine-verifiable structured artifacts