Google researchers have introduced the Biomarker Discovery Framework, a multi-agent artificial intelligence system designed to accelerate the identification of clinically meaningful biomarkers from wearable sensor data. The framework uses generative AI agents to conduct hypothesis generation, statistical analysis, and literature-grounded reasoning while maintaining strict human oversight throughout the research process. The system addresses a critical bottleneck in modern medical research: while wearable devices now generate vast streams of physiological data—from heart rate dynamics to sleep patterns—converting these signals into reliable, clinically validated biomarkers remains a manual, labor-intensive process. Existing AI-driven discovery systems often optimize for predictive performance while overlooking statistical validity, leading to spurious correlations, data leakage, and unstable features that fail to reproduce in independent datasets. Across three large-scale cohorts totaling 9,279 participant-observations, the framework autonomously identified 41 candidate digital biomarkers for mental health conditions and 25 for metabolic disease. The system combines deterministic statistical computation with generative AI reasoning, includes an adversarial validation pipeline with eleven distinct checks to prevent overfitting and confounding, and maintains human experts in the loop for all major decisions. Notably, the framework recovered known clinical signals and identified convergent biomarkers across independent datasets, demonstrating both statistical rigor and practical utility.