Agentic AI-enabled discovery across large-scale sleep physiology

Large polysomnography (PSG) archives could reveal how sleep relates to disease, but extracting insight from these complex multimodal recordings demands expert effort and defeats general-purpose AI systems. We developed AI Sleep Co-Scientist, an expert-guided environment in which human scientists direct specialist agents for hypothesis development, signal preprocessing, and statistical analysis while reviewing intermediate outputs. Every reported number is bound to the executable code that produced it, and an independent critic that cannot execute or alter the work audits each iteration. Across four cohorts comprising $\sim$124{,}000 recordings and $>$50\,TB of raw signal, we conducted five case studies and present three in detail. Diminished network-level physiological coupling during sleep was associated with incident Parkinson's disease (HR 1.48, 95\% CI 1.31--1.67) and Alzheimer's disease (HR 1.38, 1.25--1.53). Arousal dynamics characterized comorbid insomnia and sleep apnoea as an intermediate phenotype skewed towards obstructive sleep apnoea, distinguished from it by prolonged post-arousal wakefulness and more irregular arousal organization rather than greater respiratory disturbance. A comprehensive transient-oscillation scan identified a fast-sigma deficit and excess centrofrontal theta activity in narcolepsy type~1. Two further studies, on multidomain sleep ageing and short-term REM-sleep regulation, are reported in the appendix. These results show how agentic AI supports large-scale discovery when verification machinery is built around it.