Safety analysis for nuclear reactors is essential to licensing and operational assurance, yet it remains resource-intensive and dependent on multidisciplinary expertise. These analyses require the integration of diverse engineering documents, structured logical modeling, and domain-specific reasoning under strict regulatory constraints.
To support advanced reactor safety workflows, an agentic architecture leveraging generative artificial intelligence (AI) is under development at Idaho National Laboratory. The architecture decomposes safety analysis processes into specialized agents, each assigned a defined analytical role and operating under task-specific reasoning constraints. This multi-agentic structure enables the separation of data interpretation, structured safety analysis, and logical model construction into coordinated but distinct components.

Within this framework, this paper focuses primarily on agent-specific prompt engineering strategies designed to produce structured and reliable outputs aligned with safety analysis requirements. Prompts are developed to enforce hierarchical reasoning, schema-constrained generation, and consistent terminology across agents. These strategies are supported by retrieval-augmented generation (RAG) grounded in curated domain data, including existing safety artifacts and technical documentation, to improve factual alignment and reduce unsupported inference. Large language model (LLM)-generated synthetic data are also used to supplement limited domain examples and improve robustness in data-constrained safety analysis environments..

Collectively, these strategies outline a structured approach to designing agents for the integration of generative AI into nuclear safety analysis workflows.