The Advanced Research Projects Agency for Health, an agency within the U.S. Department of Health and Human Services, published a new program on January 13, 2026, with an ambitious target: building the first FDA-authorized AI agents capable of delivering round-the-clock cardiovascular specialty care without constant human supervision. Called ADVOCATE, short for Agentic AI-Enabled CardioVascular CAre TransfOrmation, the initiative represents a direct federal bet that autonomous clinical software can clear the same regulatory bar as drugs and medical devices, then prove itself in real patient settings.
What ADVOCATE Actually Proposes to Build
Most AI tools in medicine today serve as passive decision aids. A clinician asks a question; the software returns a suggestion. ADVOCATE breaks from that model. The program envisions agents that act on their own within defined boundaries, performing functions such as connecting to patient records, scheduling appointments, generating diet and physical therapy recommendations, writing or modifying prescriptions, supporting diagnoses, and escalating cases to human clinicians when risk thresholds are crossed. That list of capabilities, drawn from the ARPA-H announcement, describes something closer to a digital resident than a search engine for lab results.
The program is organized around three technical areas. The first covers the patient-facing clinical AI agent itself. The second involves a supervisory “overseer” agent designed to monitor the primary agent’s behavior and catch errors before they reach patients. The third addresses integration into existing clinical workflows, a challenge that has stalled many health-tech pilots even when the underlying algorithm works. This three-layer architecture, laid out on the ADVOCATE overview, signals that ARPA-H treats reliability and safety engineering as equal priorities alongside clinical performance.
Cardiovascular disease is a logical test bed for this kind of system. Patients often require long-term medication management, lifestyle coaching, and frequent monitoring of vital signs and lab results. Much of that work is repetitive and protocol-driven, but it still consumes scarce clinician time. An autonomous agent that could safely handle routine titration discussions, follow-up reminders, and early detection of deteriorating status could, in theory, expand access to specialty care without adding more cardiologists.
FDA Authorization as the Design Constraint
What separates ADVOCATE from the wave of health-AI startups pitching chatbots and triage tools is the explicit, repeated insistence on FDA authorization as the program’s end goal, not a future aspiration tacked on after a proof of concept. Every description of the program frames the regulatory milestone as the organizing constraint. That distinction matters because it forces performers to design for evidence standards, not just user engagement metrics.
The regulatory path, however, is far from settled. The FDA’s final guidance on clinical software draws a line between tools that merely display information for a clinician to evaluate and tools that act as regulated medical devices. An agent that autonomously modifies prescriptions or initiates scheduling based on diagnostic reasoning would almost certainly fall on the device side of that line, subjecting it to premarket review requirements. No agentic clinical AI system has cleared that process yet, which is precisely why ARPA-H describes ADVOCATE as a first-of-its-kind push.
Separately, the FDA has proposed a framework to advance the credibility of AI models used in drug and biological product submissions, outlining expectations for data quality, validation, and lifecycle management. That effort, described in an agency proposal, targets a different regulatory lane than clinical care software, but it shows the FDA is building institutional capacity to evaluate AI across multiple product categories simultaneously. ARPA-H and the FDA’s Center for Devices and Radiological Health already have a working relationship through an earlier medical imaging data partnership aimed at supporting AI and machine learning development, which could ease coordination as ADVOCATE performers approach the agency with novel submissions.
Designing for authorization from day one also means grappling with post-market expectations. Adaptive AI systems can change as they learn from new data, but regulators still need a stable description of what a product does and how its risks are controlled. ADVOCATE’s emphasis on an overseer agent and explicit escalation rules appears tailored to that concern, giving regulators defined checkpoints where human clinicians re-enter the loop.
Who ARPA-H Wants at the Table
The agency is casting a wide net for performers. ADVOCATE seeks leaders from the technology sector, academia, non-profit organizations, and small businesses to tackle the technical, regulatory, and deployment challenges the program presents. The Scalable Solutions Office within ARPA-H published the formal solicitation, designated ARPA-H-SOL-26-142, on January 13, 2026, according to federal records. A separate portal for teaming opportunities allows potential collaborators to find partners, reflecting the program’s expectation that no single organization will possess the full stack of clinical, engineering, and regulatory expertise required.
That coalition-building approach carries a risk worth watching. Programs that depend on cross-sector teams often struggle with intellectual property disputes, misaligned incentives between academic researchers and commercial developers, and slow decision-making. ARPA-H’s own operating model is still new; the agency was created to move faster than traditional federal research funding, but it has not yet accumulated a long track record of shepherding complex, regulated technologies from concept to clinic.
On the other hand, cardiovascular care is already fragmented across hospitals, outpatient clinics, device manufacturers, and digital health vendors. Any realistic deployment of an autonomous agent will have to bridge those silos. In that sense, ADVOCATE’s insistence on multi-party teams may be less a design choice than an acknowledgment of how modern health systems actually function.
ARPA-H Is Using AI to Review AI Proposals
In an unusual transparency move, ARPA-H disclosed that it is piloting secure large language model tools for the initial review of ADVOCATE proposals. A governance statement confirms that humans remain accountable for all funding decisions, with the LLM tools serving as a first-pass filter rather than a decision-maker. The disclosure is notable because it puts the agency in the position of using the same class of technology it is asking performers to build, creating a practical feedback loop and a potential credibility test. If ARPA-H cannot demonstrate responsible AI use in its own administrative processes, skeptics will question whether the clinical systems it funds can meet a higher bar.
No public data exists yet on how the LLM pilot has performed, what models are being used, or what safeguards prevent bias in proposal scoring. Those details will likely face scrutiny as the program matures, particularly from researchers and companies whose proposals are not selected. The statement emphasizes security and oversight, but without metrics on false negatives or false positives in proposal triage, outside observers will have to infer performance from who ultimately receives awards.
The Gap Between Ambition and Evidence
ADVOCATE’s central wager is that agentic AI can be made safe, reliable, and equitable enough to handle meaningful portions of cardiovascular care. Today’s evidence base does not yet support that conclusion. Most published studies on clinical AI focus on narrower tasks (image interpretation, risk prediction, or documentation assistance) rather than end-to-end management of chronic disease. Moving from point solutions to autonomous care agents multiplies the surface area for failure.
ARPA-H appears to be structuring the program to confront that gap directly. The program materials emphasize rigorous evaluation in realistic clinical environments, not just retrospective testing on curated datasets. They also call out health equity as a core objective, which implies that agents will need to perform well across diverse patient populations and care settings, including communities that have historically been underserved by specialty cardiology.
Meeting those aspirations will require more than clever modeling. It will demand careful attention to data provenance, continuous monitoring for drift, and mechanisms for patients and clinicians to contest or override agent decisions. It will also intersect with broader federal requirements around accessibility and public-facing technology. HHS guidance on Section 508 compliance underscores that digital tools procured or supported by the department must be usable by people with disabilities, a standard that autonomous clinical agents will have to meet if they are to be deployed at scale.
Legal and ethical guardrails will matter as much as technical ones. HHS web policies on disclaimers and limitations offer one model for how agencies communicate the boundaries of what their online resources can and cannot do. Translating that approach into conversational agents that feel human-like but are not human clinicians will be a delicate task, especially when patients are in distress.
For now, ADVOCATE is a funding opportunity and a design blueprint, not a deployed system. Its success will be measured not just by whether any performer secures FDA authorization, but by whether the resulting agents actually improve outcomes, reduce clinician burden, and earn the trust of patients. The program’s structure (multi-agent safety architecture, regulatory-first orientation, cross-sector teams, and even AI-assisted proposal review) suggests ARPA-H is trying to anticipate the failure modes that have plagued earlier waves of digital health hype. Whether that foresight is enough to close the gap between ambition and evidence will become clear only as the first ADVOCATE-backed agents move from code to clinic.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
