The Orchestrator
The orchestrator is the conversation-level supervisor in an Agent Stack. The router decides who handles a message. The specialist replies. The orchestrator watches the whole arc — every message, every reply, every silence — and decides whether the conversation is on track or whether something needs to change.
You can think of it as the manager looking over the specialists’ shoulders, except it never gets tired and it makes the same call every time.
What it actually does
Four things, all derived from the same state machine running over the conversation:
- Stall detection — The conversation has stopped progressing. The specialist is going in circles, or the customer keeps asking the same question. The orchestrator intervenes.
- Re-routing — The wrong specialist was picked at the start, or the topic shifted. The orchestrator hands the conversation to a more appropriate specialist mid-flow.
- Risk escalation — The customer is frustrated, threatening legal action, signaling churn, or asking for a human. The orchestrator triggers handover.
- Lifecycle tracking — Tracks where the conversation is — triaged, active with a specialist, stalled, at risk, handing off, with a human, resolved.
The full state diagram lives in the orchestrator-states reference.
What you can configure
From the Orchestrator tab, click the orchestrator node on the canvas. The slide-out lets you append routing rules to the built-in default prompt — anything you write here augments the defaults; the built-in actions and core rules always stay in effect. Use it for team-specific guardrails like:
- “For refund questions, prefer the Billing specialist over the Tech specialist.”
- “If the customer mentions delivery delays of more than 14 days, hand off immediately.”
- “Always confirm the customer’s order ID before responding about an order.”
You can also clear your saved rules to fall back to the inherited default (tenant or global), and {{TENANT_NAME}} is replaced at runtime with your team’s name.
The other orchestrator settings — model, check interval, stall threshold, risk auto-escalation, handoff mode — live in the broader stack configuration and are managed through the same panel and via global defaults.
Risk auto-escalation
When the orchestrator’s sentiment-and-risk infrastructure agent flags a conversation as risky (frustration, legal threat, churn risk, escalation request — these are the default categories, configurable per stack), the orchestrator automatically triggers handover. You don’t write rules for this; it’s the default behavior. You can disable it on a per-stack basis if you want only explicit triggers to matter.
See Infrastructure agents reference for the underlying risk-detection settings.
Re-routing without the customer noticing
If the router picks the Billing specialist but the customer’s actual question is about returns, the orchestrator can re-route to the Returns specialist mid-conversation. The customer doesn’t see the change — it just looks like the AI is helping. This is what stops a single mis-routing decision from turning into a long, frustrating thread with the wrong agent.
Hot vs. cold handover
When the orchestrator triggers handover, it can do it in two styles:
- Hot handoff — the human takes over the existing conversation immediately, continuing where the AI left off. The customer keeps typing in the same thread.
- Cold handoff — the AI prepares a handoff pack (summary, key details, sentiment, suggestions) and queues the conversation for a human. The customer is told someone will follow up.
Hot is right for synchronous channels (web chat, voice). Cold is right for asynchronous channels (email) and for outside-hours scenarios.