ARPA-H’s ADVOCATE program funds six teams to build an FDA-authorized AI agent for heart failure care

Key takeaways

ARPA-H has awarded contracts for ADVOCATE, a $62.7 million program to build an autonomous AI agent that manages heart failure patients between visits and earns FDA authorization within 24 months. Atman Health, Tempus AI and UpDoc will build the patient-facing agents, Stanford will build a supervisory layer to catch unsafe advice in real time and Johns Hopkins APL will evaluate the results independently. Duke and Kaiser Permanente will test the agents inside Epic and Oracle workflows, and ARPA-H says the model is meant to extend to any patient with any condition.

The Advanced Research Projects Agency for Health (ARPA-H) has moved its agentic AI ambitions for cardiology from concept to contracts. On September 9, the federal research agency announced the teams that will build ADVOCATE, a system meant to work around the clock as a digital member of the cardiovascular care team and to earn FDA authorization along the way. ARPA-H bills it as the world’s first effort of its kind. The program carries a $62.7 million budget over four years, and ARPA-H has committed as much as $33.7 million of it in year one.

The agent would support heart failure patients between visits, handle some care on its own, and pull in humans when the situation calls for it. The program manager, a practicing cardiologist, described the goal as “an autonomous agent smart enough to understand a patient’s treatment needs, capable of providing certain care autonomously, and equally capable of engaging the clinical team when needed.”

Why ARPA-H wants autonomous AI agents in heart failure care

According to the agency, the U.S. spends close to half a trillion dollars a year on heart disease and still trails peer nations on cardiovascular outcomes. Preventable effects of cardiovascular disease kill more than 200,000 Americans a year, and cardiologists are absent from almost half of U.S. counties.

The agency blames an approach that treats a long-term disease through occasional office visits with specialists who are in short supply. It ties that model to avoidable admissions, slow medication titration, weak adherence, and workforce burnout that lands hardest in rural and underserved communities.

ARPA-H estimates that ADVOCATE’s success would save $28 billion a year among heart failure patients alone.

Atman Health, Tempus AI, and UpDoc will build ADVOCATE’s patient-facing AI agents

ARPA-H selected three teams to build the agent patients will talk to:

• Atman Health will pair an “evidence-based clinical decision engine” with large language models and a “voice-first interface.” The system tailors its diagnostic questions to each patient as the conversation unfolds.
• Tempus AI will build on its existing Olivia patient app. The agent will watch patients continuously and dig further into the clinical picture when something meaningful shifts.
• UpDdoc will keep the conversational side of its agent apart from clinical decisions. Nothing the agent proposes goes forward until clinician-written rules confirm it fits an approved protocol.

These teams have 24 months from their contract awards to submit an FDA authorization package, and ARPA-H describes that submission as the first of its kind.

Stanford will build supervisory AI to monitor agents after deployment

Stanford University will build the watchdog layer, a “disease-agnostic” supervisory system that monitors the patient-facing agents continuously once they are live. Its job is catching unsafe advice and “out-of-distribution behavior” as it happens.

Stanford’s design runs in three stages and spends more compute at each one. ARPA-H lists them as “outlier filtering, rule-based screening, and a deep-research auditing agent.” The output is a rationale for each claim that a human can inspect.

ARPA-H has contracted with the Johns Hopkins University Applied Physics Laboratory to independently judge how well the technology works and what it does for patients. The agency also plans to work with FDA throughout the program on how this kind of AI should be regulated. The director of FDA’s Digital Health Center of Excellence called patient-facing agentic AI “one of the most consequential frontiers in digital health.”

Performers will also work with ARPA-H on common yardsticks for evaluation, on interoperability, and on how the technology gets paid for.

Duke and Kaiser Permanente will test ADVOCATE agents on Epic and Oracle EHRs

Two health system teams will design and run the implementation strategy, with an emphasis on rural and underserved communities. Duke University will validate the agents at five health systems and rural sites running Epic and Cerner/Oracle EHRs, with the American Heart Association extending the work nationally.

Kaiser Permanente’s deployment spans 21 medical centers and more than 260 clinics, with agents embedded in Epic workflows for heart failure patients. The team will run the agents in shadow mode, test them in pragmatic randomized trials, and package what it learns as blueprints built for reuse.

ARPA-H’s ambitions reach past cardiology. The agency says ADVOCATE lays the groundwork for agentic AI to join the care team for any patient with any condition. The program runs for four years, and the first FDA package is due in two.