Making waves: Beyond adoption of foundation models - charting opportunities and challenges of LLM-Agents for intelligent wastewater treatment plants

Currently, wastewater treatment plants (WWTPs) rely heavily on human-supervised control, where limited expert oversight, fragmented systems, and reactive management hinder plant-wide optimization and result in high operational costs. Large language models (LLMs) endowed with agentic capabilities (LLM-Agents), extending foundation models with perception, memory, and tool-use modules, offer a transformative avenue to address these challenges. By integrated data management, advanced tool integration and simulation, collaborative scenario analysis, reliable equipment control, LLM-Agents can coordinate the operation of existing models and control modules across the WWTPs. Acting as intelligent coordinators rather than replacements, they hold promise for adaptive and self-healing WWTPs capable of diagnosing anomalies and optimizing energy, chemical, and manpower use toward multi-objective optimization. Nevertheless, their current reliability remains limited by hallucinations and weak domain grounding. Thus, this paper outlines the boundaries and development pathways of LLM-Agents, emphasizing domain-specific customization and benchmarking as prerequisites for their responsible and trustworthy deployment in real-world wastewater management. © 2025 Elsevier Ltd.