Hypernovelty Institute Technical Intelligence
Publication posture
Technical articles carry source ledgersDaily Digest stays separateSignals become durable memorySignal Atlas adds disciplined forecasts when useful
Daily digest · June 28, 2026

The Care Loop Wants an Agent

Health AI is moving from answer boxes toward workflow authority.

Dark clinical workflow desk showing source cards, a care-loop agent gate, and an approval rail with human stop path

Lead image: health AI moving from answer boxes into reviewable workflow authority and evidence records.

The clearest current signal is ONC’s FY2026 LEAP in Health IT funding opportunity. Released June 16, it offers up to $2 million across three awards. The lead area offers up to $1 million for standards-based agentic AI in clinical care or clinical trials. ONC asks applicants to develop or adapt an agentic AI solution for a specific clinical or investigation workflow, run at least one real-world pilot, integrate with at least two health IT platforms, use FHIR-compatible standards, apply risk management, and publish an assessment report.

That is a small grant program, but the structure matters. The agency is asking for agents that can touch workflows such as patient triage, diagnosis support, care coordination, and clinical trial enrollment. It is also asking for cross-platform integration, risk assessment, and public learning. The operating question is whether a health system can prove what an AI agent was allowed to do, what data it used, what it changed, who reviewed it, and how it behaved across real software systems.

HHS has been asking the same question at the policy layer. Its RFI on accelerating AI in clinical care sought public input on regulation, reimbursement, and research and development. ONC also listed a June 25 public event to discuss takeaways from that RFI, though fuller event materials were not available in the retrieved source. The sequence still matters because reimbursement and regulation decide which tools become part of daily care. If a tool is paid for, certified, integrated, and trusted, it can become infrastructure. If the evidence layer is weak, it becomes another system clinicians have to supervise while still carrying the liability.

FDA is also pushing the proof problem into the clinical-trial workflow. In April, FDA issued an RFI on a proposed AI-enabled pilot for early-phase clinical trials. The agency said the pilot would assess how AI and data science might improve trial efficiency, safety monitoring, dose-selection decisions, and early go/no-go decisions while maintaining scientific and regulatory standards. That is a high-stakes example of the same pattern: AI can speed the decision surface, but the institution still needs evidence that the speed improved judgment rather than thinning it.

The WHO paper on AI and evidence-informed health policy widens the frame. It argues that AI is already shaping how health problems are defined, how policy options are designed, and how implementation is monitored and adjusted. Its practical recommendations include human oversight, multidisciplinary collaboration, living evidence workflows, and risk-based regulation. The public-health posture is practical: AI should support human judgment in the policy cycle instead of replacing it.

FDA’s device guidance points in the same direction. Its lifecycle draft for AI-enabled device software functions focuses on documentation and risk management across the total product life cycle. Its guidance on predetermined change control plans asks sponsors to describe planned modifications, the methodology for developing and validating them, and the expected impact. In plain English: if the model changes after launch, the change needs a governed path.

Health-care AI is resolving into accountable care loops: data access, workflow integration, model change, patient safety, reimbursement, clinician review, and public trust.

Verification bottleneck

Verification is becoming the scarce institutional function.

  • Clinical workflows can move faster than medical records, reimbursement rules, and oversight committees can verify agent authority.
  • AI-enabled trials can accelerate safety monitoring and early decisions, but sponsors and regulators still have to verify data quality, decision quality, and patient protections.
  • Model updates create a live evidence problem. Someone has to verify whether the system that was approved or purchased is still the system being used.
  • Watch next: whether health AI pilots publish usable assessment reports or leave hospitals, clinicians, patients, and payers with vendor claims they cannot inspect.

Opportunities

Where value may appear: health AI workflow proof packets.

Someone could build practical tools or services for clinics, small health systems, trial sites, digital-health teams, patient advocates, and procurement groups: agent authority maps, FHIR integration checklists, clinical workflow evidence folders, model-change logs, patient-facing explanation sheets, risk-review templates, and pilot assessment summaries that translate technical claims into reviewable operating records.

This is idea fodder only, not medical, legal, regulatory, cybersecurity, procurement, financial, or investment advice. The useful operator test is direct: before the agent touches care, show the record of what it may do, what it did, and who can stop or correct it.

Sources