Towards autonomous urban drainage modelling: evaluating AI agent architectures for automated SWMM calibration

Abstract

Urban drainage model calibration requires coordinating data preparation, parameter screening, simulation, validation and interpretation, and remains slow and expertise-intensive. AI agents, which use large language models to plan and execute multi-step tasks, could automate it, yet how an agent should be organised is unknown. Using a controlled Storm Water Management Model (SWMM) calibration case study, we compared three organisations of the same procedures: written instructions consulted when needed (skill-based), fixed interfaces supplied in advance (tool-based), and specialised agents each responsible for one stage (multi-agent). All three completed the workflow and produced calibrated models of equivalent accuracy, so they differed in the reliability and cost of reaching that result. Consulting instructions on demand sustained longer sequences of requests but required a more capable language model, whereas supplying interfaces in advance was least expensive and suited single tasks but became unreliable once the task was extended. Dividing the work among agents remained reliable at a cost that varied with the language model. No single organisation was best on all counts, so the choice depends on the length of the task, the capability of the available language model and the acceptable token consumption. These complementary strengths indicate where each organisation is appropriate, and a complete calibration workflow could combine them, each applied to the stage it suits, offering a practical route towards agent-supported urban drainage modelling.