The Feed: June 29, 2026
HHS Wants AI Agents to Diagnose and Prescribe. The Foundation They'll Run on Doesn't Exist Yet.
ARPA-H is building an FDA-authorized autonomous AI agent for cardiovascular care, and the data layer the agent would have to read hasn’t been built.
The FDA just dropped its warning against Whoop’s blood pressure feature, and the wearables behind Whoop are next into your flowsheet.
KLAS says ambient speech, smart phrases, and order sets drove EHR satisfaction gains for physicians; nurses fell further behind because the builder programs were never for them.
Patient messages to providers are up 153% across 2,000 Epic hospitals, office visits are up too, and the AI drafting tools health systems are deploying are the faster horse.
Eighty percent of cancer patients would rather hear the diagnosis from their physician, but the configuration treats it as just another lab result.
Off-Label: A health-economics podcast spent 27 minutes explaining what prior auths actually are. AI is about to learn to copy those decisions.
The Feed is a quick take on a few of the highest-profile health IT news items of the week, with one piece from outside the beat at the end. Free through June 2026, then paid.
Last week ARPA-H announced ADVOCATE, the first FDA-authorized agentic AI system for cardiovascular care. The program manager called it the moment AI moves from assisting clinicians to acting on its own. Federal health officials want autonomous AI agents that diagnose patients, triage them, and change or prescribe medications without routing every decision back to a human clinician.
It’s the wrong fight. The data the agents will read isn’t trustworthy yet. No one is asking what is doing the supervising.
Start with the data layer. The cardiovascular grouper at most Epic shops is a mess of overlapping, contradictory definitions built one project at a time and never reconciled. The AI agent has to pick one to act on. Most organizations don’t know which one their agent will pick. Most don’t know that the choice matters. That problem doesn’t get solved by adding an FDA authorization on top.
The supervisor question is harder. Tune the supervisor agent too tight and you flood the escalation queue with alerts no human can clear. Tune it too loose and the liability spikes. The only realistic architecture is a deterministic stack that evaluates real-time risk tiers rather than static binary thresholds, which is the kind of system almost no one is building because it requires giving up the comfort of a single threshold rule. The shortcut everyone will take is to put another LLM on top of the action agent. Now you have AI grading AI.
That’s the deepest problem. Much of the data the supervisor agent will treat as ground truth was generated by upstream AI in the first place. Ambient AI wrote the note. AI coding tools picked the diagnosis. AI-drafted replies populated the messaging history. The supervisor reads all of that and treats it as the patient’s clinical reality. Call this AI cannibalism. The model trains on a corpus the model helped write, then grades itself against the corpus it polluted. ADVOCATE is the federal endorsement of building the second floor before the foundation is poured.
The FDA Just Cleared The Wearable Laundromat. Garmin And Fitbit Are Next.
The FDA withdrew its 2025 warning against Whoop’s blood pressure tracking feature after Whoop argued the feature supports general wellness rather than medical use. The agency agreed. The data keeps flowing.
Last week’s column argued the Epic flowsheet doesn’t know where the data came from. The field just records the number, and the clinician who signs the chart owns the diagnostic liability. The FDA just removed the regulatory friction that might have slowed that problem from playing out at scale.
The architectural question gets harder from here. Most health systems are ingesting wearable data through patient-entered flowsheets, which means a downstream human has to manually validate the volume of incoming data points. A single Whoop device can send hundreds of heart rate or pulse ox readings per hour. Once iWatch, Garmin, Fitbit, and the rest start filtering in, are they all going to flow into the same flowsheet row, or will each device get its own ecosystem? If it’s the latter, every organization is about to build a different data ecosystem per device, one quiet implementation decision at a time, with no one consolidating the policy. That sounds like a mess.
The Builder Program Was For Physicians. The Documentation Rolls Downhill To Nurses.
The KLAS Arch Collaborative 2026 report found most organizations improved over the past year, driven by ambient speech, smart phrases, and expanded order sets. Nurses fell further behind, with lower satisfaction tied to duplicate documentation and usability gaps.
The trade press is reading this as a story about what’s working. The real story is who the work was built for. The cutting-edge tools have all been directed at physicians. In the Epic world, the physician-builder concept has traction. The nurse-builder concept does not. Epic modeled a clinical-builder program on the physician-builder structure to close the gap, but it has never landed with the same status or attention.
Part of the gap is staffing. Many Epic analysts come from nursing backgrounds, so there’s an assumption that base is covered. The reality is that nurses who become analysts tend to evolve into pure analysts and lose their frontline nursing mindset. The other reality is structural: documentation rolls downhill, and the people at the bottom of the hill are nurses. Every workflow optimization that saves a physician keystrokes adds a flag, a flowsheet field, or a confirmation step for the nurse who follows them. The KLAS report is the receipt.
AI Drafted Replies Are The Faster Horse. The Triage Layer Underneath Is The Redesign.
A new JAMA study analyzing Epic EHR data across more than 2,000 hospitals and 47,000 clinics found patient-written portal messages rose from 0.99 to 2.5 messages per patient per year between 2020 and 2025. Clinician-authored messages rose 24% in the same window. Office visits also rose. The messaging is additive, not substitutive.
The reading the industry is going to take is that messaging is up and we need better tools to draft replies. That’s the faster horse. AI-drafted-reply tools are part of the solution, but the meta layer is the actual opportunity, and almost no one is working on this. The redesign is an ingestion mechanism for the physician, the office manager, or the entire group that uses an LLM to surface which messages need manual review and pushes the bulk of the rest into automation. Health IT is still wedded to deterministic, if-this-then-that expert systems. The available primitives changed years ago. The muscle memory has not.
Epic’s Emmie platform is the build behind the digital concierge from the last issue. It’s trying to address the patient-side of this by wrapping context and security around a ChatGPT-style conversation experience. The output is trailing the user experience ChatGPT defined. However, it is new and I am confident Emmie will catch up. The triage layer on the clinician side is the harder build, and it’s the one that would actually reduce workflow volume rather than help clinicians keep up with it.
The Cures Act Made Transparency Mandatory. Nobody Built A Way To Make It Humane.
A survey covered in HIStalk’s Monday Morning Update found 80% of cancer patients would prefer their physician deliver the diagnosis rather than read it through a patient portal release. Seven percent learned of their cancer diagnosis through the portal. Half of those were alone when they read it.
Health IT has been dealing with this question since 21st Century Cures made results notification immediate and mandatory. The finding isn’t surprising. The response that matters is: how would you actually execute on it? If there are certain results a patient wants to see immediately and other results they want delivered by a physician, there is no good lever to deliver the logic. The discrete data is not available to make the distinction. Some systems have built coarse, deterministic rules based on CT scans, pathology, or specific lab orders. None of those rules look at the qualitative aspect of the result. The patient who learned of their cancer diagnosis through the portal got that result because the configuration treated it as just another lab.
Same ceiling as the messaging story above. Deterministic rule engines hit their limit at the point where qualitative judgment starts, and most Epic shops configured the release rules to the regulatory minimum and called it done.
Throughline
Five stories this week. An announcement of autonomous AI agents that will diagnose and prescribe. A regulatory reversal that makes ingesting consumer wearable data easier. A report on EHR satisfaction. A peer-reviewed study on patient messaging. A survey on how cancer patients want to learn they have cancer.
The throughline isn’t AI. The throughline is that every one of these stories is downstream of a data layer most organizations haven’t built yet, governed by rule engines that were designed for a deterministic world and are still being asked to handle qualitative work. ADVOCATE will not work if the cardiovascular groupers underneath it aren’t governed. The flowsheet absorbing Whoop data will not produce trustworthy signal if it can’t tell where the number came from. The AI drafting in-basket replies will not solve volume if the triage layer underneath stays deterministic. The cancer portal release will not become humane until the system can read more than the lab type.
Federal health officials want AI agents that act on their own. That is the headline. The foundation those agents would have to act on is still mostly absent.
Off-Label
Off-Label. Not health IT, not my usual lane, but something that made me think. Here’s why it should matter to people who do what we do.
Stacey Richter spent 27 minutes this week on her podcast explaining what a prior authorization actually is. Most people think a prior auth is a doctor’s office calling an insurance company to make sure the patient really needs a drug. It isn’t. It’s a bargaining tool. The middlemen who decide which drugs get covered use prior auths to punish the cheaper drug and protect the deals they’ve cut with the more expensive one. The patient is barely in the room. The first federal pilots are now training AI to approve or deny prior auths the way the humans do. The AI is learning to copy a decision that was never about whether the patient needed the drug. The contamination started before the AI showed up.
John Lee is an emergency physician and Epic consultant who helps health systems bridge the gap between Epic’s capabilities and operational reality. He specializes in data architecture, registry optimization, and making Epic’s tools actually deliver results.
If you need help configuring your Epic environment to support these capabilities, connect with him on LinkedIn (https://www.linkedin.com/in/johnleecmio/) or via his website (https://www.hitpeakadvisors.com/).