Large language models (LLMs) have made remarkable progress in recent years, demonstrating impressive capabilities across a wide range of general tasks. However, their effective application in vertical domains remains limited. Accordingly, this study systematically explored the application method, evaluation framework, and capacity enhancement strategies for LLM agents in assisting hydropower operation. We defined three representative tasks for the LLM-based agent, including hydropower generation scheduling, reservoir water level control, and station operational data analysis. Based on these three seed tasks, this study selected the Dadu River Basin as a case study and generated 300 agent testing tasks. Furthermore, a four-stage execution assessment framework encompassing task recognition, data querying, core tools calling, and results integration was established to evaluate the agent’s task performance. Results indicate that current LLM-based agents exhibit significant deficiencies in comprehensively identifying required data for task completion. To address this shortcoming, this study applied context engineering to automatically enrich background information and provided a data reference checklist to facilitate agent understanding of necessary data. This improvement increases agents’ execution success rate during the data acquisition stage from an average of 52.9% to 81.3%, and increases the end-to-end execution success rate from 4.3% to 46%. This study demonstrates the practical application of LLM-based agents in assisting hydropower operations and provides insights for future LLM-based agents’ capability to better execute hydropower operation tasks.