Article ·
MARA Demo Update: Earned Authority and a Person on Call
Six changes to the MARA demo, each one an answer to something published in October 2026.
Contents9 sections
What changed
The MARA demo has new screens. MARA, the Medical Agentic Resident Assistant, is a design for an assistant that keeps one patient's whole record. Each person in that patient's care sees the part their role needs. The demo walks one fictional patient, Dana Whitfield, from the triage call for a hurt ankle to the insurer's claim decision. It runs entirely in the browser, and no real patient sits behind it.
This update adds six things, and every one of them came from reading what clinicians, regulators and researchers published this month. Open the interactive demo and follow along; the MARA project page has the full design. The terms the piece leans on are in the glossary at the end.
Authority that is earned, and can be lost
On 13 October 2026 the American Medical Association (AMA) published a governance framework for agentic AI that takes on clinical functions. Its central principle is that clinical authority "should be earned through evidence and maintained through demonstrated performance." It asks four questions of every such system: what function the AI performs, what shows it can perform that function and what happens if it fails, what it should be permitted to do and under what conditions, and who is responsible for the risks and outcomes.
MARA already had a competency ladder, the five steps a skill climbs as it proves itself: Heard, Tried, Held, Trusted and Autonomous. The demo now shows it where the decision is made. A strip headed "MARA's authority here" sits on the doctor's, pharmacist's, specialist's, coder's and radiology screens, and it names each function's step. On the doctor's screen, the pre-visit brief is Trusted, step 4 of 5, and drafts for the provider to review. Visit-note drafts and treatment answers sit a step lower at Held. There MARA drafts or advises, and Dr. Raman signs or decides.
A skill climbs only on measured outcomes, and the demo now shows the ladder working in the other direction too. In the learning loop, a card from an earlier patient records a demotion. On 30 September a six-week outcome showed MARA's "back to full duty in 2 weeks" estimate was wrong for a similar patient, who was re-injured. Return-to-work estimates went from Trusted to Held, step 4 to step 3, and that skill now drafts only while the provider signs. That demotion sits in the clinic's AI governance log rather than in this visit's audit, since it happened on someone else's case.
Authority that can only go up isn't earned.
Graded on the whole decision
OncologyBench, a benchmark from Latch Bio, ran AI agents through 25 multistep cancer-care and cancer-research evaluations, with 31 combinations of model and agent framework each trying every evaluation three times. Across those 2,325 attempts, 57% passed every sub-task. Only 19% also got the final decision right, which is one attempt in three among those whose every step was correct. Correct steps didn't add up to a correct decision.
A second study shows what clinicians actually ask. A JAMA study led by Mass General Brigham looked at about 271 million queries that 952,008 US clinicians put to OpenEvidence, a clinical search assistant, between January 2024 and June 2026. Among physicians, treatment and management questions made up 33.1% of queries and medication questions 14.1%, against 13.9% for diagnosis. That's more than twice as many "what do I do" questions as "what is it" questions. Three coauthors worked for OpenEvidence during the study, one of them as its chief medical officer, and the paper discloses it.
So the doctor's screen now carries a treatment question. Dr. Raman asks, "Ibuprofen or acetaminophen for her pain, given her med list and her sulfa allergy?" MARA answers with its evidence graded as FACT, BELIEF or HYPOTHESIS, each line with its source, and her authority badge reads "Held · advises only — provider decides."
MARA's answer reads the patient's current medicine list. Add sertraline, an antidepressant that raises bleeding risk alongside ibuprofen, to Dana's list at intake on the pre-evaluation nurse's screen. Now the recommendation switches to acetaminophen, 1 g up to three times a day, and says why ibuprofen was withheld. Dr. Raman can accept it or change it. Her decision, which may differ from MARA's suggestion, is what the pain line in the visit note follows. The learning loop records whether she accepted or changed it, so the 2-week check-in can grade both. The whole recommendation is what gets judged, end to end.
A person, any time
Google published its AMIE study in The Lancet on 8 October, run with Beth Israel Deaconess Medical Center. Before urgent care visits at the hospital's primary care clinic, 98 patients talked with AMIE, a research diagnostic AI. Supervising physicians watched each conversation in real time against safety criteria set in advance, and not one conversation had to be stopped. What MARA takes from it is the study's shape: the AI was supervised live, and a clinician stayed in charge.
In the demo, the patient app now has a "Talk to a person" card, and the transfer is there at any point. Pressing it hands the conversation to a named nurse, Maya Chen, RN, on the nurse advice line, who picks up within 15 minutes. Her reason and the transcript go with it, so Dana doesn't have to start over. Even the handoff is written to the audit log.
Records the patient brings, and an audit that names who allowed it
In mid-October HHS announced its 2026 LEAP in Health IT awards, the Leading Edge Acceleration Projects grants run by the Office of the National Coordinator for Health IT (ONC). MedStar Health Research Institute won one for a project called Sovereign Execution: Vigilant, Gated and Verifiable Agentic Care. It pairs patient-controlled data storage, built on the Solid protocol, with cryptographic verification and tamper-evident certificates. Every route can be checked.
Two demo changes follow from that. First, a record the patient uploads is labelled patient-supplied. Dana's export from MyChart, her own patient portal, shows as "Unverified until matched" until it's tied to the organization that sent it. It sits apart from records that arrived over the national network, which come signed by their source.
Second, the compliance log has two more columns. "Allowed by · authority" names who allowed each AI action and at what step. For the triage summary it reads "Maya Chen, RN · Trusted · drafts — nurse signs." "Chain (illustrative)" gives each entry a sequence number and a short marker tied to the entry before.
That marker is illustrative only. The page recomputes it from the list on screen, so it draws a hash chain's outline and detects no tampering. MARA as designed stores each entry's full hash, taken over every field, and verifies it against a trusted anchor, so a changed or deleted entry breaks the chain.
Senators Josh Hawley and Chris Murphy announced the bipartisan AI Agent Accountability Act on 1 October. As the sponsors describe it, an operator whose agent recklessly causes hacking damage would face civil and criminal liability under the Computer Fraud and Abuse Act (CFAA). Developers would be liable too, for failing to build reasonable safeguards against hacking they knew, or had reason to know, the agent could do. No bill text had been published when this was written. Whatever happens to it, an operator will need to show what each agent was allowed to do and who allowed it, and that's what the new audit columns are for.
What the demo is
MARA is a design, and the demo touches no real patient. Dana is fictional, the data is invented, and MARA's responses are scripted to show how the screens behave. Its citations come from general knowledge and would be checked before any real use. In every screen a person makes the decision.
Try it
- The interactive demo. Start with the guided tour, then open the doctor's screen, the learning loop, compliance and audit, and the patient app.
- The MARA project page, for the full design.
The next question on Dana's chart is the 2-week check-in, and it will grade Dr. Raman's choice and MARA's suggestion in the same row.
Sources
- American Medical Association, "A governance framework for agentic AI assuming clinical functions", 13 October 2026, checked 20 October 2026.
- Latch Bio, "OncologyBench: Benchmarking AI Decisions in Cancer Research and Treatment", 13 October 2026, checked 20 October 2026.
- Patel et al., "Clinician Use of an Artificial Intelligence–Enabled Clinical Information Platform", JAMA, 15 October 2026, with the Mass General Brigham release for the coauthors' OpenEvidence roles, checked 20 October 2026.
- Google, "AMIE clinical study in The Lancet", 8 October 2026, checked 20 October 2026.
- Healthcare IT News, "HHS awards $1.5M for agentic AI and laboratory interoperability projects", 14 October 2026, for the 2026 LEAP in Health IT awards, checked 20 October 2026.
- BERI, "AI Agent Accountability Act targets the company running the agent", October 2026, relaying the sponsors' statements of 1 October 2026, checked 20 October 2026.
- The MARA demo,
demos/mara-demo.html, as carried to the site by pull request 249 of the site repository on 20 October 2026. Figures 2 to 5 were captured from that version.
Glossary
| Term | Definition |
|---|---|
| Agentic AI | Agentic AI is software that takes steps toward a goal on its own, such as looking things up, drafting a note or routing a request, rather than only answering one question. The AMA framework and the demo both deal with that kind of system. |
| AMA | The American Medical Association is the largest US physicians' organization. Its October 2026 framework sets out how agentic AI should earn and keep clinical authority. |
| AMIE | AMIE is Google's research AI for diagnostic conversations. Its Lancet study ran it with live physician supervision before urgent care visits. |
| Audit log | The audit log is the record of every event in a visit, with the rule that applied and who decided. In the demo it now also names who allowed each AI action. |
| Authority strip | The authority strip is the row headed "MARA's authority here" on each clinical and coding screen. It shows each function's step on the competency ladder and what that step permits. |
| CFAA | The Computer Fraud and Abuse Act is the main US federal law against computer hacking. The Hawley and Murphy bill would extend its liability to the operators and developers of AI agents. |
| Competency ladder | The competency ladder is the five steps a MARA skill climbs on measured outcomes: Heard, Tried, Held, Trusted and Autonomous. A later outcome that proves a recommendation wrong moves the skill down a step. |
| FACT, BELIEF, HYPOTHESIS | These are the grades MARA puts on each line of evidence. A fact has a checked source, a belief has support that stops short of settled, and a hypothesis is still waiting to be validated. |
| Hash chain | A hash chain links each log entry to the one before it through a fingerprint of its contents. Changing or deleting an entry breaks every link after it, so tampering shows. |
| LEAP in Health IT | The Leading Edge Acceleration Projects are grants from ONC for new health IT. One 2026 award funds MedStar's work on patient-controlled, verifiable AI routing. |
| MyChart | MyChart is a widely used patient portal. In the demo, Dana uploads her own MyChart export, and MARA holds it as unverified until it's matched. |
| ONC | The Office of the National Coordinator for Health IT is the part of HHS, the US Department of Health and Human Services, that sets health IT policy and runs the LEAP awards. |
| OncologyBench | OncologyBench is Latch Bio's benchmark of 25 multistep cancer-care and cancer-research evaluations. It scores an AI attempt as passed only when every step and the final decision are right. |
| OpenEvidence | OpenEvidence is an AI search assistant that clinicians use to look up medical evidence. A JAMA study counted what about 952,000 clinicians asked it. |
| Sertraline | Sertraline is a common antidepressant from the SSRI family. Taken with ibuprofen it raises the risk of bleeding, which is why the demo's treatment answer changes when it's on the list. |
| Solid | Solid is an open protocol for keeping personal data in stores the person controls and granting apps access to it. MedStar's LEAP project builds patient-controlled storage on it. |
| Trusted anchor | A trusted anchor is a fixed, independently held reference that a hash chain is checked against. Without one, someone who rewrites the log could rewrite the chain too. |
Table 1. The terms this piece leans on, each in plain words.
Keep reading
More field notes
This piece is part of the MAX Research Collective library. Browse the rest, or connect on LinkedIn.