ONC Funds Open-Source Guardrails for Clinical AI Agents

The US Office of the National Coordinator for Health IT (ONC) has put roughly $1 million behind an effort to build open-source guardrails for AI agents that carry out clinical follow-up work, in what is one of the first federal grants aimed specifically at agentic AI in patient care. Announced on October 6, the award to MedStar Health Research Institute is small by industry standards, but it targets a question hospitals are already asking: how do you let software act on a patient's behalf over weeks or months, and prove afterwards that it did what it was supposed to?

What ONC announced

According to ONC's announcement, two awards totalling $1.5 million were made under the 2026 LEAP in Health IT program, which has funded interoperability work since 2018. The notice of funding opportunity was posted on June 16, 2026 and closed to applications on July 16. It named three areas of interest: agentic AI in clinical care, expanding ONC's Lantern API-monitoring tool with community feedback, and laboratory interoperability. ONC says it received applications only for the first and third areas, so awards were made in just those two.

Trade outlet Health System CIO reports the split: MedStar Health Research Institute received $999,985 for the agentic AI project, and the Regenstrief Institute received $500,000 for a laboratory terminology project. ONC's own announcement does not list individual amounts, so those figures rest on the trade report.

The MedStar project

MedStar's project is titled "Sovereign Execution: Vigilant, Gated and Verifiable Agentic Care." Per ONC, it aims to give patients control over the AI that executes their care by building the Sovereign Execution Bridge, described as open-source middleware that coordinates complex clinical follow-ups. Three technical ideas stand out:

• Long-running processes. Follow-ups such as chasing a test result or confirming a referral unfold over days or weeks, so the system is built around durable workflows rather than a single chat exchange.
• Patient-gated data pods. Data sits in patient-controlled stores built on the Solid protocol, so access is granted by the patient rather than assumed by the system.
• Verifiable routing. AI routing decisions carry tamper-evident certificates, meaning a reviewer could check afterwards how a task was routed.

ONC's project page lays out three priority areas: developing a national taxonomy of clinical follow-ups and configuring the Bridge for it; running shadow-mode human-factors review, which it calls the "Training Wheels" phase, benchmarked against finite-state-machine, directed-acyclic-graph and order-based baselines; and validating the system with real patients in a clinical pilot where humans approve actions. The page calls the Bridge a "fully open-source middleware harness" but gives no licence details.

Plans, not results

It is important to be clear about what exists today. Everything above describes planned work. Health System CIO stresses that the capabilities are planned rather than proven, and nothing in ONC's materials reports test results, error rates or patient outcomes. The grant funds development and evaluation; it does not show that agents can safely manage clinical follow-ups.

The timeline is also reported inconsistently. ONC's project page lists an estimated 2026–2028 period, while Health System CIO says both projects run through September 2029, citing ONC spokesperson Cannon Leavelle. We have not been able to reconcile the two, so the schedule should be treated as uncertain.

Why it matters for health systems

Oversight is the bottleneck

Much of the commercial excitement around agents in healthcare concerns administrative work such as scheduling and documentation. Clinical follow-up is different because the agent's actions can affect care, which raises the bar for audit trails and human sign-off. A design that logs routing decisions in a verifiable way and keeps humans in the loop speaks to those concerns directly.

An open reference point

Because the code is meant to be open source, Health System CIO suggests it could serve as a reference for evaluating the controls offered by electronic health record and platform vendors, and that its test results could help health systems decide what to build themselves versus buy. That is the publication's view of the potential, not an outcome of the grant.

How practitioners are thinking about it

The same report quotes a September interview with Ryan Smith of Intermountain Health, who said platform vendors keep adding agents inside their own products and that no single vendor provides a full enterprise solution for agents working across systems. Intermountain is building what it calls an agentic control plane over roughly 12 months, and for its IT service-ticket agents uses separate models for quality and safety testing, which Smith described as "AI watching AI, with people watching both." Health System CIO notes his comments concern a different use case and predate the ONC announcement, so they are context rather than evidence about MedStar's approach.

The second award

Regenstrief's project is not about AI agents, but it is part of the same announcement. Per ONC, the institute plans to recruit at least three small, independent laboratories and guide each through readiness assessment, data quality assessment, targeted terminology fixes, validation and long-term sustainability. Steven Posnack, ONC's principal deputy national coordinator, said laboratory data is among the most frequently exchanged and clinically consequential in healthcare, so targeted improvements can have "an outsized impact on patient care." Clean, standardised lab data is also a precondition for any AI system that reasons over results, which is why the two awards sit comfortably together.

Wider federal context

Health System CIO also points to an ARPA-H award announced on September 9 that supports Kaiser Permanente's evaluation of agentic AI in heart failure care, with cardiologist oversight built into the design. We did not open the underlying announcement for that award, so we rely on the trade report for its description. Taken together, the two awards suggest federal funders are choosing to test agent designs that keep clinicians in control before backing broad deployment.

What to watch next

• Early artefacts. Whether MedStar publishes the follow-up taxonomy or an initial code release, and under which licence.
• Shadow-mode findings. Any reported comparison against the workflow baselines named in the project, and what the human-factors review finds about where agents go wrong.
• Pilot design. How the clinical pilot defines human approval gates, and which follow-up types are included first.
• Schedule clarity. A resolution of the 2028 versus 2029 completion dates.

The practical takeaway is modest but real: before agents are trusted with clinical follow-up, funders want evidence that their actions can be gated by patients, checked afterwards and examined by researchers in the open. This grant is a start on that evidence, not the evidence itself.