Hi all - I wrote up a pattern for letting AI agents write to FHIR safely — instead of the agent writing the intended resources directly, it drafts a change proposal: a persisted transaction Bundle (updates as FHIRPath Patch ops) that a human reviews and approves before it executes. Think Terraform plan/apply for clinical writes. It digs into the stuff that only bites after the happy path: drift while a draft sits in the queue, version-pin re-basing, validation tiers, reviewer automation bias. Curious if others here are building similar HITL review flows for agentic proposals — would love to compare notes! https://www.osbytes.io/blog/change-proposals-in-fhir-hitl-review-for-ai-driven-writes