The benchmark signal
On September 9, 2026, ARPA-H announced selected teams for its Agentic AI-Enabled Cardiovascular Care Transformation program. The program aims to develop a reliable, FDA-authorized clinical agentic AI system for cardiovascular care. ARPA-H's public structure includes patient-facing AI, a supervisory agentic layer, clinical validation, and a scalable deployment pathway rather than treating model performance as the entire program.
Deployment is being funded as part of the architecture.
ARPA-H's award directory describes one deployment team planning enterprise-scale implementation across 21 medical centers and more than 260 clinics, while another consortium is designed to test agents across five health systems and rural sites using both Epic and Cerner/Oracle environments. Those are planned program activities, not proof of successful clinical deployment.
Why this is worth studying
Assemble Teams analysis: consequential AI programs should be reviewed as socio-technical deployment systems. A clinically impressive model can still fail the mission if supervisory controls, workflow integration, validation cohorts, escalation paths, EHR integration, regulatory evidence, patient trust, and operating ownership are weak.
ADVOCATE's public materials also emphasize measurable safety, outcomes, and equity, and describe a clinician-extender model where AI may act autonomously in some circumstances while engaging the clinical team as needed. That makes authority boundaries and escalation design part of the mission—not post-launch policy paperwork.
A five-layer deployment architecture
What can the agent observe, ask, recommend, or act on?
Define the intended use, user population, modality, accessibility needs, and what the system must never do without escalation.
Who or what challenges the agent?
Separate generation from clinical authority where appropriate. Define safety rules, anomaly detection, escalation, and the conditions requiring clinician involvement.
What population and workflow evidence is needed before scale?
Plan for multi-site testing, subgroup performance, workflow variation, failure analysis, and evidence strong enough for the intended regulatory pathway.
Can the system operate inside real care infrastructure?
EHR integration, identity, data quality, clinical messaging, documentation, scheduling, monitoring, and local governance are product dependencies.
Who owns the system after deployment?
Define monitoring, incident response, model or rule updates, clinician training, patient support, change control, and accountability for performance drift.
Why this complements Open Directory
Open Directory can surface health systems, AI developers, validation partners, researchers, and missions working on clinical AI. A Mission Intelligence note should show how those actors fit into an execution architecture: who builds the patient layer, who supervises, who validates, who integrates, who operates, and which evidence must exist before claims of readiness become credible.
What the evidence does not prove
ADVOCATE is an active R&D program. Announced awards and planned deployments do not establish clinical benefit, regulatory authorization, cost savings, safety, or successful scale. ARPA-H's projected outcomes are program goals, not realized results. This note therefore treats the program as an execution architecture worth studying—not as proof that agentic AI is ready for broad autonomous clinical use.
Program review question
If the model were technically excellent tomorrow, which non-model dependency would still prevent safe deployment at a real health system? That dependency belongs in the product plan now.
Editorial standard: facts trace to ARPA-H public materials; planned activities are labeled as planned; Assemble Teams interpretation is labeled. This note does not imply ARPA-H, HHS, FDA, awardee, or health-system endorsement.