For the commercialization of marine renewable energy, a systematic decision-support framework is required to evaluate power potential and economic feasibility by integratively considering marine environmental conditions and wind resources at the early development stage. In particular, floating offshore wind projects involve large-scale, heterogeneous metocean datasets and multiple high-fidelity analysis models, making automated data management and workflow control critical challenges. This study proposes an LLM-based AI agent–driven data architecture for site selection and techno-economic assessment of floating offshore wind systems in Korean waters. Long-term wave and wind conditions are generated using the SWAN (Simulating WAves Nearshore) numerical model and processed into structured metocean datasets. These data are normalized and used as inputs to validated in-house power performance models to estimate annual energy production, while capital and operational expenditures are evaluated based on empirical data and existing literature and linked to levelized cost of energy analysis. The AI agent is implemented using the Qwen3 large language model, not as a predictive or inferential model, but as an intelligent orchestrator that directly invokes and coordinates high-accuracy in-house analysis codes. This approach enables efficient handling of large metocean datasets while preserving the reliability of physics- and data-driven models. The proposed architecture integrates environmental data, energy production assessment, and economic evaluation into a consistent data flow, providing a scalable foundation for automated site selection and early-stage decision support for floating offshore wind development.