AI that acts — schedules, files, submits — is governed here the way clinicians have been governed for a century: a defined scope of action, a proctoring period, action-level audit, and automatic suspension the moment an agent acts outside its scope. Every transition writes to the same tamper-evident chain as the models. The framework reference is public at /agent-governance/.
| Agent | Status | Permitted actions | Actions | Breaches | Proctoring ends | |
|---|---|---|---|---|---|---|
| Referral triage router composite vendor · autonomous |
Suspended | classify_referral route_to_queue |
2 | 1 | 2026-09-28 | |
| Prior-authorization submission agent in-house · supervised |
Privileged | draft_pa_request submit_pa_request check_pa_status |
4 | 0 | 2026-09-28 | |
| Ambient documentation scribe Meridian Scribe (composite) · supervised |
Proctoring | draft_encounter_note attach_note_to_encounter flag_low_confidence_section |
4 | 0 | 2026-10-28 |
Three agents, three states — the whole framework on one screen. The scribe is in its proctoring period (every action reviewed). The prior-auth agent graduated: a clean proctoring record, so a named human granted privileges. The referral router was suspended automatically the moment a vendor update made it act outside its registered scope — no meeting required, and the entire sequence is on the tamper-evident chain in the overview feed. The public framework: /agent-governance/.