The estimation of design floods in small watersheds is crucial for flood control and disaster mitigation. However, conventional methods suffer from drawbacks such as tedious parameter extraction, complex calculation processes, and a low degree of automation. To address these challenges, this paper presents an automated approach for design flood estimation in small watersheds based on a multi-modal large language model agent. The methodology first involves constructing a multi-modal parameter set for design floods, which integrates spatially interpolated grids with digitized charts. Subsequently, an intelligent agent is developed with DeepSeek-V3 as its core, supported by a modular API framework that provides services for data access, the Rational Method, and the Guangdong Provincial Comprehensive Unit Hydrograph method. Leveraging a tool-calling mechanism, the agent can autonomously interpret natural language instructions from users to plan and execute the entire workflow for design flood estimation. A case study on the Pajiang River Basin demonstrates that the agent can independently manage the full process, from task parsing to final report generation. The results show a high degree of consistency with those obtained from the official Guangdong Hydrological and Water Resources Design and Calculation Software Platform. This proposed method offers a novel technical path for automating complex hydrological calculations and validates the effectiveness and accuracy of multi-modal LLM agents in the field of smart water conservancy.