Summary
Provides consensus guidance and evaluation methods for healthcare AI agents that can plan and take actions across systems, addressing autonomy boundaries, permissions, identity and tool access, human verification, provenance, audit logging, escalation, failure modes, and ongoing monitoring.
Healthcare Implications
Healthcare organizations and developers should define which actions an agent may take, apply least-privilege access, preserve traceable records of inputs and actions, require human verification for consequential steps, test foreseeable failure modes, and establish escalation and monitoring processes before deployment.