Most of the debate about regulating AI in medicine is about catching up: how a framework built for static devices handles software that learns and acts on its own. ARPA-H, the federal government’s health-research agency, is running at the problem from the other direction. With a program called ADVOCATE, it is deliberately trying to force the precedent into existence, by funding a piece of autonomous clinical AI engineered to land inside the FDA’s authority and make the agency rule on it.
What the program is actually doing
ADVOCATE, the Agentic AI-Enabled Cardiovascular Care Transformation, aims to produce the first FDA-authorized agentic AI for clinical care. The distinction matters. To date, no clinical agentic AI has been FDA-authorized; the AI the agency has cleared is predictive, tools that flag, score, or estimate, with a clinician acting on the output. Agentic AI is different in kind. It can plan and execute a sequence of tasks on its own. No authorization pathway exists for that in a high-risk clinical setting, and ADVOCATE’s premise is that the way to get one is to build a system that demands it.
The program funds two pieces. The first is a patient-facing agent for cardiovascular disease, the country’s leading cause of death and, by the CDC’s longstanding estimate, the source of roughly 200,000 preventable deaths a year despite treatments that are often inexpensive and widely available. The second is the more telling one: a supervisory “overseer” agent designed to monitor deployed agents for continued safety and efficacy. That second system targets a genuine gap in healthcare AI, the lack of ongoing monitoring for systems that keep learning after deployment, and it is meant to be disease-agnostic, so it can be reused across conditions later.
The disease choice is the strategy
The most revealing decision in ADVOCATE is why it starts with the heart, and the regulatory logic is the brief’s reading of it: cardiovascular disease is high-risk and well-defined, which puts the resulting device clearly inside the FDA’s authority rather than in the grey zone where many lower-risk tools quietly operate. Pick a condition unambiguous enough that the FDA cannot decline to engage, and you produce a precedent that can then extend to other conditions. The disease is the wedge.
The program’s own leader frames the choice in compatible but more clinical terms. Haider Warraich, a cardiologist who advised the FDA commissioner before joining ARPA-H, has said the appeal of heart disease is that it is highly treatable with mostly generic, inexpensive medications and lifestyle changes, and that few conditions carry as rich a base of clinical-trial evidence, so the problem to solve is scale and access rather than medical knowledge. He describes the goal as a clinician-extender, an autonomous agent that can support patients directly and pull in the clinical team when needed, and he is direct about the regulatory gap: no clinical agentic AI has been authorized yet, and ARPA-H means to work with the FDA to build the pathway.
What it signals, and what it does not
ADVOCATE is a notable move because it inverts the usual posture toward AI regulation in medicine. Instead of waiting for autonomous tools to arrive and then asking how to oversee them, the government is commissioning one designed to be overseen, and using a deliberately hard case to build the rulebook.
What it does not do is settle anything yet. This is a funding program with a multi-year horizon: teams selected around mid-2026, phased development, a down-select that eliminates approaches along the way, and large-scale studies of clinical outcomes, safety, cost, and, crucially, reimbursement before anything reaches authorization. Each of those stages can change the answer. And the reimbursement question waiting in the later phases is the same one that decides whether any cleared clinical tool actually reaches patients: a precedent for authorizing autonomous AI is not the same as a path to paying for it.
Still, the intent is the story. The bottleneck for autonomous clinical AI was never only the technology; it was the absence of a regulatory pathway anyone had walked. ADVOCATE is the government deciding to walk it on purpose, in the hardest place it could pick, so that the precedent it sets is one regulators cannot wave away.
The Counter Brief — one email, every Monday.
The week's AI-for-revenue moves in a 5-minute read: which tools are worth the budget and which to skip, plus what to do this week. Source-checked, no vendor decks.
Edited by Aditya Marin Gasga
Free. One click to unsubscribe.
Frequently asked questions
What is ADVOCATE?
The Agentic AI-Enabled Cardiovascular Care Transformation (ADVOCATE) is an ARPA-H program aiming to develop the first FDA-authorized agentic AI for clinical care, autonomous AI that can plan and execute clinical tasks, not just make predictions. It funds two systems: a patient-facing cardiovascular-care agent and a supervisory 'overseer' agent that monitors deployed agents for ongoing safety.
Why start with cardiovascular disease?
Two reasons. As a matter of regulation, cardiovascular disease is high-risk and well-defined, which places the technology squarely inside the FDA's authority rather than in the grey zone where many lower-risk tools sit. As a matter of medicine, the program's leader, cardiologist and former FDA adviser Haider Warraich, points out that heart disease is treatable with mostly generic drugs and has an unusually rich clinical-evidence base, so the bottleneck is scale and access, not knowledge.
How is this different from existing healthcare AI?
To date, no clinical agentic AI has been FDA-authorized. The AI the FDA has cleared is predictive: tools that flag, score, or estimate, with a clinician acting on the output. ADVOCATE targets generative, agentic AI that autonomously plans and acts in a high-risk clinical setting, a category with no existing authorization pathway. The program is explicitly trying to create that pathway.
What is the timeline?
ARPA-H planned to select teams of innovators within roughly six months of the early-2026 launch, around mid-2026, followed by a multi-phase development and 'down-select' process. Later phases include large-scale studies of clinical outcomes, safety, cost-efficiency, and reimbursement, on a roughly three-year horizon to FDA authorization.