Weekly signal

Agentic AI for healthcare continued its move from lab demos to live workflows this week. Key signals: a major health system trialing intake agents; frontier multi-agent models and "Work" agents from OpenAI that make long-running, tool-enabled workflows practical; increased standards and security attention from U.S. agencies and the research community; and new benchmarks and community workshops to evaluate agentic medical behaviors.

What changed

1. Penn Medicine began live use/testing of K Health’s agentic patient‑intake technology inside its virtual primary care/OnDemand service — the agent collects history, triages simple needs (e.g., refills), and produces structured summaries that feed clinicians’ workflows, with human clinicians retaining final authority. This deployment is positioned as an expansion point if results show time‑savings and better visit quality.

2. OpenAI launched GPT‑5.6 (July 9) and ChatGPT Work (July 9), which together add multi‑agent orchestration modes, longer-running scheduled tasks, stronger programmatic tool use, and desktop/browser computer‑use features — capabilities that materially lower the engineering effort to build agentic workflows that touch EHRs, scheduling, and admin tasks. OpenAI also describes layered safeguards and enterprise controls for organized deployments.

3. Standards and security work is advancing: NIST published a summary analysis of responses to its RFI on AI‑agent security (recently available) and federal agencies continue OneHHS/ONC work to accelerate clinical AI adoption while emphasizing governance and risk management — showing policy attention shifting from models to agent behaviors and supply‑chain risk.

4. Research infrastructure is maturing: HealthAgentBench (a realistic benchmark suite for agentic healthcare tasks) and an Agentic AI for Medicine workshop (MICCAI community) are active, making it easier to evaluate long‑horizon agent performance and safety in clinically realistic scenarios.

What to do with it

• If you run clinical operations: pilot intake/triage agents in low‑risk pathways (refill, administrative triage) with clear opt‑out routes, logging, and human‑in‑the‑loop gates; measure time saved and downstream clinician satisfaction.
• If you build clinical agents: evaluate GPT‑5.6/Work for multi‑step orchestration but instrument programmatic tool calls, auto‑review, and access controls; plan for per‑workflow risk assessments aligned with NIST/ONC guidance.
• If you set policy or procurement: update vendor questionnaires to ask about long‑running agent behaviors, parallel agent orchestration, audit trails, and retention policies; track HealthAgentBench and MICCAI workshop outputs as emerging evaluation norms.

Sources: see numbered list below.