Making waves: Human-governed workflow logic for agentic AI in water management

Agentic AI is extending artificial intelligence from content generation to task execution, enabling large language models to coordinate knowledge, data, models, and tools in water management. As these systems enter real operations, ensuring task process quality by balancing reliability and efficiency becomes increasingly important. Two linked needs therefore emerge: grounding agent execution in complete operational logic and avoiding repeated reconstruction of procedures that have already been verified. In this Perspective, we argue that execution trajectories provide a natural bridge between the two. We advocate a human-governed learning loop in which automated and expert evaluation distils recurrent, reliable trajectories into reusable skills, while workflow controllers learn from or retrieve these skills to guide subsequent tasks. By enabling operational experience to progressively and organically evolve into reusable planning and execution capability while preserving domain rules, traceability, and professional accountability, this loop could support the broader adoption of Agentic AI in water management.