We’ve gotten used to AI that predicts risks or flags anomalies, but we haven’t yet trusted it to act. That line is about to blur. In a move that signals a major shift from “doctor’s assistant” to “autonomous provider,” the federal government is launching an ambitious program to put AI agents directly in charge of patient care tasks—and they want the FDA to sign off on it within three years.
The News
On January 21, 2026, the Advanced Research Projects Agency for Health (ARPA-H) announced the “ADVOCATE” program, soliciting proposals to build “agentic” AI systems for cardiovascular care. Unlike today’s predictive tools, these generative AI agents are designed to independently schedule appointments, adjust medication dosages, and guide patient diet and exercise. The initiative, described as a “big swing” by the Trump administration, aims to secure FDA authorization for these autonomous tools in just 39 months.
The program, led by cardiologist and former FDA official Dr. Haider Warraich, focuses specifically on heart failure—a condition contributing to 200,000 deaths annually. Crucially, the project isn’t just building one bot. It will fund the creation of two distinct systems: a clinical agent to treat patients, and a “supervisory agent” designed to monitor the clinical AI’s performance in real-time. The plan is to co-develop these tools with health systems and deploy them locally for testing after two years.
Why It Matters
This is a regulatory moonshot. To date, the FDA has largely cleared AI that acts as a second set of eyes for clinicians. Approving an “agentic” system—one that can execute tasks without a human pressing every button—would establish a massive precedent for the entire industry. Warraich sees this project as a “forcing function” to compel regulators to figure out how to oversee generative AI. “The FDA cannot authorize an AI doctor,” Warraich noted, “but they can authorize a technology that improves outcomes for patients with heart failure.”
The introduction of the “supervisory agent” is perhaps the most significant technical innovation. One of the biggest barriers to deploying autonomous AI is the “black box” problem—we don’t always know why models make decisions, or when they might drift. By funding a separate, disease-agnostic watchdog AI, ARPA-H is attempting to solve the problem of ongoing monitoring for continuously learning systems. If successful, this dual-agent architecture could become the standard safety harness for all future medical AI.
For patients, the stakes are tangible. Heart failure management relies on constant tweaking of diuretics and other meds. An always-on AI agent could prevent the subtle, unmonitored declines that lead to hospitalization, scaling specialty care in a way that the human workforce simply cannot match.
The Skeptic’s View
The safety implications here are profound, and the medical community is divided. The American Heart Association (AHA) has expressed support for the initiative, though they were careful to clarify that they do not view AI as a replacement for cardiologists.
However, the American Medical Association (AMA) is far more cautious. CEO John Whyte warned that delivering cardiovascular care without clear physician oversight raises “serious concerns,” noting that “heart care is complex, and mistakes in diagnosis or treatment can be life-threatening.” A medication adjustment that makes sense by the numbers might be dangerous in the context of a patient’s other unrecorded symptoms. There is also a tension between the proprietary nature of the clinical agents and the need for public trust; while Warraich hopes the supervisory agent will be open source, the clinical tools likely won’t be.
Looking Ahead
The clock is ticking. ARPA-H plans to select innovator teams by June 2026. Innovators will have a year to develop their prototypes before a “down select” process weeds out the weaker proposals. Watch closely to see if tech giants partner with major academic medical centers to bid for this contract. Whoever wins won’t just be building a product; they’ll be writing the rulebook for the next decade of digital health regulation.