Multi-agent AI for end-to-end air quality analysis and decision support

Urban air pollution poses increasing risks to public health and environmental sustainability, motivating more automated, data-driven management. Large Language Models (LLMs) offer promise for such automation, yet their deployment in atmospheric governance remains constrained by limited access to real-time, heterogeneous environmental data and insufficient cross-disciplinary reasoning for comprehensive pollution analysis. Here we present Neuro-Air, an end-to-end multi-agent LLM framework that couples real-time air pollution data retrieval with a persistent code-execution sandbox. Built on a hierarchical architecture, a central coordinating agent directs five domain-specialized worker agents, emulating the thorough workflow of human experts in urban air pollution management. Across 317 real-world cases, Neuro-Air executed the full pipeline from heterogeneous data ingestion to structured output. In a matched ablation across five additional backbones, the multi-agent configuration achieved a pooled 88.4% accuracy on a common 90-task programmatically verified subset, scored against ground truth recomputed from the source data. Expert review further delineates an operating boundary, separating tasks where autonomous analysis is dependable from those requiring human judgement. Rather than a one-stop solution, Neuro-Air is positioned as a decision-making assistant that pairs reliable automation with clear human oversight, offering an early foundation for applying LLMs to real-world air-quality management.