Noah-Agent: A Multi-Expert AI Agent Framework for Automated Parameterization and Validation of Large-Scale Fortran Climate Models

Land surface models such as Noah-MP routinely exceed 100,000 lines of Fortran code, creat- ing prohibitive onboarding barriers for observational scientists. While recent AI coding agents succeed in general software engineering, they cannot sustain the end-to-end modeling activities required in scientific computing—modifying numerical schemes while preserving conservation laws, recompiling with appropriate flags, and validating against observational benchmarks. We present Noah-Agent, a multi-expert AI framework addressing three critical gaps: the translation gap (Fortran represents only 0.04% of GitHub code, limiting LLM training data), the integration gap (fragmented custom protocols), and the validation gap (absence of physics- aware validation harnesses). Noah-Agent makes three contributions. First, automated Fortran climate model genera- tion handling the full workflow from knowledge retrieval through compilation and validation. Second, the first application of the Model Context Protocol (MCP) to atmospheric modeling, es- tablishing reusable patterns for the climate community. Third, a multi-expert architecture with specialized agents—Model Cartographer, Configuration Steward, Code Synthesizer, and Vali- dation Marshal—that decompose complex tasks into manageable subtasks with physics-aware validation gates. Our methodology integrates proven agent design patterns with domain-specific tools, positioning Noah-Agent as a template for WRF, CESM, and other community models.