ARPA-H, the U.S. government’s health “moonshot” agency, plans to commit $62.7 million to develop partially autonomous AI that would help treat heart failure — and, crucially, be authorized by the FDA to do so. It is one of the most ambitious bets yet on putting artificial intelligence in a genuinely clinical, decision-making role.

What ARPA-H is proposing

The program, called ADVOCATE, aims to produce the first FDA-authorized “agentic” AI — software that doesn’t just flag information for a human but can take actions within defined guardrails. The vision is AI agents trusted to provide 24/7 specialty-level support: assessing symptom severity, ordering lab tests, adjusting medications, and modifying appointments, diet and exercise plans. The first year commits $33.7 million, with the remainder subject to renegotiation.

The problem it targets

Heart failure is the deadliest chronic disease in the United States, affecting roughly 6.7 million Americans. A central, frustrating reality is that effective, guideline-recommended treatments exist — but many patients never get optimally treated because access to cardiologists is limited, especially in rural and underserved areas. Doses go un-titrated, warning signs go unnoticed, and patients cycle in and out of the hospital. ARPA-H’s bet is that a tireless AI agent could help close that specialist gap, catching deterioration early and keeping therapy optimized between visits.

Who’s building it

Initial awards went to health-technology firms Atman Health, UpDoc and Tempus AI, plus academic and health-system teams from Stanford, Duke and Kaiser Permanente — a mix of startups and established institutions.

Why it matters — and the hard questions

The ambition is striking: autonomous software that can prescribe and order tests is a significant leap beyond the “AI as assistant” tools common today, and it raises real questions about safety, liability and trust. That is precisely why the FDA-authorization requirement is the linchpin — it is the guardrail meant to ensure such a system is rigorously validated before it touches patient care. If it works, it could be a template for AI in other chronic diseases; if it stumbles, it will be a cautionary tale about the limits of automating medical judgment. Either way, ADVOCATE is a notable marker of where clinical AI is heading. Not medical advice.