Multi-Agent AI Health Agent Outperforms Single-Agent Systems

A multi-agent artificial intelligence (AI) system that reasons over wearable and medical data received higher ratings than single-agent and parallel multi-agent systems from both end users and health experts, according to a study published in Nature. The authors are from Google Research, Google DeepMind, and collaborating institutions.

Through analysis of consumer health queries, user surveys, and an expert workshop, the investigators identified four user needs: general health knowledge, personal data interpretation, wellness advice, and symptom assessment. They then built the personal health agent (PHA), an open-source framework that runs on the Gemini 2 family of large language models (LLMs). In the PHA, an orchestrator coordinates three specialist agents: a data science agent, a health domain expert agent, and a health coach agent. Evaluations used data from the Wearables for Metabolic Health (WEAR-ME) study (n = 1,165) and involved more than 7,000 annotations and 1,350 hours of effort across 11 benchmark tasks.

In live multi-turn conversations, 45 end users ranked the PHA as their top choice for overall preference 46.7% of the time. They chose the parallel multi-agent system 28.9% of the time and the single-agent system 24.4% of the time. Three health experts rated overall quality at 72.8% for the PHA, compared with 64.3% and 54.4% for the two baselines, respectively.

Among the specialist agents, the data science agent passed 75.5% of code tests on the first attempt, compared with 58.4% for the base Gemini model. The domain expert agent answered 83.6% of more than 2,000 expert-level multiple-choice questions correctly, compared with 81.8% for the base model. The questions came from board certification and coaching examinations. In an expert safety review of health summaries generated by the domain expert agent, 83% had no or minimal potential harm, 15% had mild harm, and 2% had moderate-to-severe harm.

The authors described the PHA as a research prototype for consumer wellness, not a diagnostic or clinical triage system, and said it should not replace professional care. They noted that the evaluation measured conversational quality and short-term task efficacy, not behavioral change or longitudinal health outcomes. The evaluation also relied partly on LLM-based autoraters and was Fitbit-centric. The PHA averaged 6.5 LLM calls and 205.3 seconds per response round, compared with 3.0 calls and 36.4 seconds for the single-agent system.

Source: Heydari AA, Gu K, Srinivas V, et al. The anatomy of a personal health agent. Nature. Published online October 21, 2026. doi:10.1038/s41586-026-11161-8