Water-MAS: A multi-agent LLM framework with instruction-data decoupling for smart water management

Sustainable water management demands transforming heterogeneous, long-sequence water data into timely, data-driven insights. While single-agent systems can perform basic tasks through tool integration, they suffer from low accuracy, limited adaptability, and insufficient big data analysis capabilities, rendering them unsuitable for complex, multi-step water management workflows. To overcome these constraints, we propose Water-MAS, a large language model (LLM)-based multi-agent framework designed for diverse water management tasks. The framework integrates three specialized agents (planning, execution, and checking) and introduces an instruction-data decoupling mechanism that separates task instructions from the underlying data, reducing computational overhead. Experiments demonstrate that, in four-tool invocation tasks, multi-agent collaboration achieves 89% higher recall (vs. single-agent baseline) under instruction-data decoupling, and the decoupling mechanism boosts recall by 143% relative to the coupled approach within the multi-agent framework. Analysis across foundation models reveals that smaller-scale models match performance of larger counterparts (Qwen3-8B vs. GPT-OSS-20B: F1 scores = 0.68 vs. 0.72), providing guidance for balancing system performance and model size. Case studies on water sampling and water quality forecasting validate the framework's real-world applicability. This work offers a reusable framework for deploying LLM-based multi-agent systems to enable automated tasks in environmental monitoring and data analysis.