Coastal zone management faces significant challenges from heterogeneous multi-source data integration, complex numerical modeling workflows, and the semantic gap between scientific requirements and executable analytical pipelines. We present GeoCoast-Agent, an LLM-driven multi-agent framework that translates scientists’ natural-language requirements into executable coastal scenario-analysis workflows. GeoCoast-Agent coordinates specialized agents for task understanding, workflow planning, model execution, evaluation, and reporting through a controlled library of parameterized scientific skills. Each skill is defined by typed inputs, allowable parameters, applicability conditions, execution interfaces, and validation rules. Scientific queries are converted into constrained skill instances and mapped to predefined workflow templates, which are subsequently adapted through parameter selection and evaluation feedback. A case study in Xiamen Bay illustrates the framework’s ability to coordinate heterogeneous data and modeling tools across an end-to-end scenario-analysis process. GeoCoast-Agent provides a reproducible and extensible approach to scenario analysis within coastal digital twins while lowering the technical barrier to interdisciplinary coastal research.