ARPA-H Bets $62.7 Million on Agentic Heart-Failure AI

On Sept. 9, 2026, ARPA-H named the first teams in a four-year $62.7 million program to build FDA-authorized agentic AI for heart-failure care. UpDoc, Tempus AI and Atman Health will each try to field a patient-facing agent that works between visits.

The wager is narrower than the “AI doctor” talk around it. All three teams keep the language model on the phone, and put clinical authority in a rules system a physician can inspect.

ARPA-H Puts $62.7 Million on a Digital Teammate

The Advanced Research Projects Agency for Health calls the effort ADVOCATE, for Agentic AI-Enabled Cardiovascular Care Transformation. Year one is funded at up to $33.7 million. If the work holds, the agency estimates $28 billion a year in savings across the heart-failure population alone.

Haider Warraich, M.D., the program manager and a practicing cardiologist, framed the gap as access, not chemistry. “Every day, Americans are dying from cardiovascular disease that we know how to prevent and treat,” he said. The barrier, he added, is a clinician who can guide the care.

What we are trying to achieve with ADVOCATE is a technology that can serve as a clinician-extender: 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.

Haider Warraich, M.D., ADVOCATE program manager, ARPA-H

ARPA-H says the United States spends nearly half a trillion dollars a year on heart disease and still posts worse results than peer countries. More than 200,000 Americans die each year from preventable effects of the disease. The stated product is a 24/7 digital care teammate that can change appointments, medicines, diet and exercise, then hand the case back to a human when it should.

THE OPENING STAKE

• Program size: Up to $62.7 million over four years, with $33.7 million in year one.
• FDA clock: Patient-facing teams must file a first-of-its-kind authorization package within 24 months of award.
• Claimed payoff: $28 billion a year in heart-failure cost savings if the system works.
• Independent check: Johns Hopkins University Applied Physics Laboratory will score technical performance and clinical outcomes.

Sharif Vakili, M.D., CEO of Palo Alto-based UpDoc and a primary-care physician at Stanford Medicine, said medicine has always been limited by how many clinicians exist and how many hours they can give patients. UpDoc’s slice of the award is up to $9.2 million, with Microsoft, OpenAI and NVIDIA listed as supporters of that team. Tempus and Atman did not publish their dollar figures.

The Counties That Never Got a Cardiologist

The geography is the reason a research agency is writing this check. Warraich is not only the program manager. He is also a co-author of a 2024 research letter in the Journal of the American College of Cardiology that mapped the holes in the map.

That letter found 1,454 counties had no cardiologist, or 46.3% of all 3,143 U.S. counties, covering 22 million people. Rural counties were almost empty of the specialty: 86.2% had none. Counties that did have cardiologists averaged 24 of them.

WHAT THE 2024 LETTER MEASURED

• The drive: Round-trip distance to the nearest cardiologist was 87.1 miles in counties without one, against 16.3 miles in counties with one.
• The risk: Counties without a cardiologist had a 31% higher cardiovascular risk index and higher age-adjusted heart death rates.
• The years: Life expectancy was one year shorter in the counties that had no specialist.

Heart-failure drugs are not experimental. Atman Health, the Boston team in the patient-facing group, notes that four medicine classes have to be started and raised over months, each step checked against kidney function, other labs and blood pressure. In major registries, it said, only about 1 percent of eligible patients were on target doses of at least three of those classes.

Guideline-directed therapy already exists. The stubborn problem is getting a patient onto those doses and keeping them there when the next titration visit is an 87-mile round trip. A talking agent does not fill a pharmacy order, and it cannot make someone stay on a beta blocker that makes their legs swell. That is the objection clinicians have raised since ADVOCATE was first described in January, and it is the test the trials will have to meet.

Three Teams Share One Safety Split

Technical Area 1 is the patient-facing agent: software that supports people with heart failure between visits and escalates to the human team. ARPA-H picked three performers. Their marketing differs. Their architecture does not, at least on the point that matters for a regulator.

THE PATIENT-FACING TEAMS

Rahul Patel, Atman’s CEO, said the unfashionable choice was the point. Encoding cardiology as explicit rules is slower than training a large language model. He said they did it on purpose.

An LLM cannot be given permission to act autonomously on reasoning nobody can examine, and permission is the whole game here.

Rahul Patel, co-founder and CEO, Atman Health

Rahul Deo, M.D., Ph.D., Atman’s chief medical officer, put the clinical risk in one sentence. “In heart failure, a single wrong adjustment can put a patient in the emergency department with elevated potassium, kidney injury or an arrhythmia,” he said. Before a system can make those calls, a physician has to be able to ask what was checked, why, and whether it gets the same answer every time.

Tempus, the public company in the group, is extending Olivia rather than starting from a blank agent. The firm Olivia personal health concierge app went national on Jan. 21, 2025, as a place to pull records, device data and a smart profile, not as a prescriber. ADVOCATE is the step that would let that product do clinical work when a patient’s status actually changes.

UpDoc Already Cleared a Narrower Insulin Agent

UpDoc is the only TA1 team that already has an FDA-cleared patient-facing agent, and the clearance is tighter than the company’s later branding. The UpDoc insulin 510(k) record, K253281, was received on Sept. 29, 2025 and cleared on Dec. 23, 2025 as a Class II drug-dose calculator.

The indication is medication management for adults with type 2 diabetes. The software gives insulin plan instructions from a clinician-specified plan. Patients log glucose, meals, symptoms and adherence by tap, voice or text, or from a Bluetooth meter or continuous glucose monitor. A predetermined change control plan was authorized with the clearance.

That is a talking calculator under a physician’s plan, not a general clinician. UpDoc said as much in its ADVOCATE statement: the current insulin system works inside prescribed plans and defined bounds. The heart-failure architecture, in which larger models would reason over changing data and pick diagnostic and treatment actions inside UpDoc’s Clinical Intelligence System, “has not been cleared by the FDA.”

Thomas Osborne, chief medical officer at Microsoft Federal, said too many patients still wait too long for specialty heart care. The company’s role, as listed by UpDoc, is support, not a separate award. OpenAI and NVIDIA are in the same supporting line.

Stanford Builds the Watchdog Layer

Technical Area 2 is the admission that TA1 can fail in the wild. Stanford University won the supervisory-agent contract. After a patient-facing system is deployed, this layer is supposed to watch it in real time for unsafe advice and for behavior that sits outside the training distribution.

Stanford’s design is a three-stage, compute-escalating pipeline: outlier filtering, rule-based screening, and a deep-research auditing agent that writes an inspectable rationale for each claim. ARPA-H wants that overseer to be disease-agnostic, so it can follow whatever TA1 product survives the down-selects.

Rick Abramson, M.D., director of the FDA’s Digital Health Center of Excellence, tied the program to a regulatory job the agency does not yet have on the shelf. He said getting the framework right takes close, repeated work among developers, clinicians and the FDA, and called ADVOCATE the kind of effort required for continuous monitoring of autonomous clinical AI for safety and effectiveness.

Warraich, speaking in UpDoc’s announcement, said ARPA-H is working with the FDA and the Office of the National Coordinator for Health IT, plus health systems, medical societies, payers and technology firms. The bet is that the rules, the payment path and the software can be built together, instead of shipping a model and hoping the system moves.

A Five-System Trial on Two Record Platforms

Technical Area 3 is where an agent either becomes clinic software or stays a demo. Two health-system teams will run implementation, with a stated focus on rural and underserved sites.

WHO RUNS THE BEDSIDE TEST

Kaiser’s plan is the closest thing in the roster to a national operations manual: embed the agent in Epic for heart-failure patients, watch it in shadow mode, then compare agent-enabled care with usual care in pragmatic trials. Duke’s job is messier on purpose. Two record vendors and rural sites are where an always-on agent either interoperates or dies in an interface queue.

UpDoc said performers that clear competitive milestone reviews get a funded path from prototype to clinic, including verification studies, head-to-head scoring against cardiologists, and trials under an FDA Investigational Device Exemption. Later phases are not guaranteed. Continued work depends on those reviews.

What the 24-Month FDA Package Has to Prove

ARPA-H told TA1 teams to submit a first-of-its-kind FDA-authorization package within 24 months of contract award. Atman published the tightest private timetable: FDA pre-submission in year one, an Investigational Device Exemption filing at month 12, a single-center shadow-mode study, and a De Novo submission targeted inside 24 months.

THE FILING CLOCK ATMAN PUBLISHED

1. Year one: FDA pre-submission talks on an agent that acts between visits.
2. Month 12: Investigational Device Exemption application, so trials can generate the evidence package.
3. Before De Novo: A single-center shadow-mode study, with the agent running beside usual care.
4. Inside 24 months: De Novo submission, the path for a device that has no predicate of this kind.

UpDoc’s insulin tool used a 510(k) against a dose-calculator predicate. A heart-failure agent that adjusts several drug classes, diet and exercise, then writes into the record, is a different ask. De Novo is the honest label for that. Shared evaluation standards, interoperability rules and reimbursement paths are listed as program work, not as extras.

WHAT WE KNOW

• The roster: Atman, Tempus and UpDoc in TA1; Stanford in TA2; Duke and Kaiser in TA3; JHU/APL as the outside scorer.
• The money: $62.7 million over four years, $33.7 million in year one; UpDoc disclosed up to $9.2 million.
• The gate: A TA1 FDA-authorization package is due within 24 months of award.

WHAT IS UNCONFIRMED

• Other awards: Tempus and Atman have not stated their contract values.
• Who advances: Later phases depend on competitive milestone reviews; no team is guaranteed the full four years.
• The product: No heart-failure agent in this program is FDA-authorized, and UpDoc says its next architecture is not cleared.

Vinod Khosla, whose firm backs Atman, repeated a line he has used for a decade: a village should be able to get better cardiac care than a top academic center, but only if the same evidence is applied the same way every time, and someone can check when it is not. That is the standard the 24-month package has to meet, in counties that still do not have a cardiologist to do the checking in person.

Disclaimer: This article is news reporting on a federal research program and company statements. It is informational only and is not medical advice, a treatment recommendation, or an investment recommendation. Readers should not change medicines, devices or care plans based on this reporting, and should consult a licensed physician or other qualified clinician for personal medical decisions and a licensed financial adviser before acting on any company or contract news. Figures, team lists and regulatory timelines reflect the ARPA-H, FDA, company and journal sources cited as of Sept. 9, 2026, and may change as awards, filings and trial results move.