Orchestrator States Reference
Every conversation handled by an Agent Stack moves through a state machine controlled by the orchestrator. The state determines what happens next — whether the specialist replies, whether re-routing kicks in, whether handover triggers.
You can’t directly edit these states. They’re driven by what’s happening in the conversation. Knowing them helps you understand what the orchestrator is doing and why.
The states
- triaged — A new message arrived. The router is about to pick a specialist
- active_specialist — A specialist has been assigned and is handling the conversation
- stalled — The conversation has stopped progressing — repeated turns without resolution
- re_triage — The orchestrator decided the wrong specialist was picked; routing again
- risk — Sentiment-and-risk infrastructure agent flagged the conversation as risky
- handover_pending — Handover has been triggered; prerequisites are being collected
- human_active — A human has picked up; AI is no longer replying
- resolved — The conversation is done — by the AI or by a human
Transitions
- (start) —
triaged— New customer message arrives triaged—active_specialist— Router picks a specialist successfullyactive_specialist—stalled— Stall threshold hit — too many turns without progressactive_specialist—resolved— Specialist marks the issue resolvedactive_specialist—risk— Sentiment / risk detection firesstalled—re_triage— Orchestrator decides a different specialist would helpstalled—risk— Stall persists past the escalation thresholdre_triage—active_specialist— New specialist assignedrisk—handover_pending— Handover triggered; collecting prerequisiteshandover_pending—human_active— Prerequisites complete, conversation queued to a humanhuman_active—resolved— Human resolves and marks Done
What you can tune
- Stall threshold — how many unprogressed turns before
active_specialistbecomesstalled. Lower = more aggressive intervention. Higher = the AI gets more chances. - Risk auto-escalation — when the sentiment-and-risk agent flags a conversation as risky, transition to
riskautomatically. Default on. - Check interval — how often the orchestrator evaluates state. Shorter = faster intervention, more compute. Longer = cheaper, slower to react.
These settings live on the orchestrator config and inherit from tenant or global defaults if you haven’t customized per stack.
What’s the same as conversation status
Don’t conflate orchestrator state with the conversation status. The conversation status is the inbox lifecycle (Active / Snoozed / Done / Archived) — visible to your team in the conversation list. The orchestrator state is the AI-internal state of an in-progress conversation. A conversation with orchestrator state human_active will typically have a conversation status of Active in the inbox.