Agentic Workflow Architecture for Environmental Remote Sensing Analytics

Environmental remote sensing analysis requires complex workflows and domain expertise that many potential users lack. We present Terra AI, an agentic system in which a large language model orchestrates remote sensing tools and machine learning services, translating natural-language queries into executable multi-step workflows. The system integrates Google Earth Engine operations with independently deployed ML models — an algal bloom classifier and a peat moisture estimator — exposed as Model Context Protocol (MCP) servers. Each MCP server carries its own prompt instructions that encode domain-specific workflow rules, while a core system prompt provides general orchestration guidance. We evaluate orchestration reliability using a benchmark adapted from TaskBench, comparing a minimal and an optimal prompt configuration across 20 test cases. The optimal configuration improves tool selection F1 from 0.71 to 0.89 and tool ordering F1 from 0.79 to 0.99. Results indicate that explicit workflow rules are the dominant factor in reliable tool chaining, and that independently developed ML models can be made accessible to nonspecialists through standardized tool interfaces.