Exploring a multi-agent framework with large language models for nuclear design automation

Traditional nuclear reactor design workflows involve complex simulations, parameter optimization, and result analyses, and many of these steps still rely heavily on expert manual effort, which limits efficiency, reproducibility, and scalable engineering deployment. To address this problem, we developed MANDiF-CORCA, a multi-agent framework that couples large language models (LLMs) with an industrial deterministic reactor core analysis code for auditable workflow automation. The framework combines artifact-gated execution, state-machine-constrained coordination, and structured code-based data collection so that each workflow stage advances only after the required intermediate files are successfully generated and verified. Using an HPR1000-based reactor model, we evaluated four LLM backbones on nine representative tasks and further demonstrated the framework on six engineering scenarios, including critical boron search, distributed power analysis, control rod worth evaluation, reactivity-coefficient calculation, and operational-domain assessment. In the comparative evaluation, average workflow success rates ranged from 83.54% to 96.30%, with DeepSeek-V3.2-Exp achieving the best overall performance. Robustness tests further showed that the framework remained stable for specific tasks but became more sensitive to model temperature and semantic perturbation in more complex workflows. Overall, the study shows that coupling LLM agents with industrial deterministic reactor design software provides a feasible path toward robust, auditable, and reproducible workflow automation in nuclear design.