Weekly signal

This week (covering developments visible between September 21–29, 2026) healthcare moved from discussion to concrete agent deployment activity: health systems are building EHR-embedded agents, federal research dollars are funding autonomous clinical agents for cardiology, major vendors are shipping nurse-facing agent features, and governance warnings about "shadow" agent deployments intensified.

What changed

1. Epic hosted its second Agent Factory Build‑A‑Thon where 25 health systems prototyped EHR‑embedded AI agents for concrete clinical and operational tasks (transfer triage, medication refill checks, day‑of‑surgery cancellation risk, neonatal risk‑flagging and more). Epic says training for early adopters starts in October with broader availability in 2027.

2. ARPA‑H’s ADVOCATE program (Agentic AI‑Enabled CardioVascular CAre TransfOrmation) advanced to awards and public plans: three TA1 performer teams (including Tempus, Atman Health, UpDoc) will build patient‑facing autonomous cardiac agents while Stanford, Duke, Kaiser Permanente and Johns Hopkins APL are lined up for supervision, validation and enterprise deployment designs; ARPA‑H frames this as the first effort to build an FDA‑authorizable class of autonomous clinical agents.

3. Oracle Health expanded its Clinical AI Agent to nurses—embedding voice navigation, AI summaries and real‑time dictation inside the inpatient EHR—signaling vendor moves to put agentic features directly in chart workflows rather than as bolt‑ons.

4. Governance alarm: Imprivata’s new survey found 72% of U.S. healthcare organizations report AI tools or agents are being deployed without formal IT approval ("shadow AI"); the report urges identity‑led governance and action controls to prevent unauthorized access or agent actions. Coverage amplified safety concerns about agent autonomy and auditability.

5. Underlying platform enablement continued: OpenAI announced an Epic EHR integration for ChatGPT for Healthcare to bring authorized chart context into clinician workflows—an enabler for agent experiences that must be governed inside the EHR.

What to do with it

• If you run clinical/IT programs: treat agentic pilots as integration projects—require shadow‑mode evaluation, clear human‑in‑the‑loop gating, and EHR‑level audit trails before any write‑capability.
• Security teams: prioritize identity, least‑privilege and agent action control (Imprivata playbook concepts) and log all agent activity for rapid rollback.
• Builders and vendors: prefer in‑EHR embedding or vetted connectors, design supervisory/monitoring agents, and instrument explainable rationales for every high‑risk action (ADVOCATE’s supervisory architecture is a template).
• Executives: budget for training, formal change controls, and liability/BAA reviews before enabling agents with write access to records.