Skip to main content
A conversation flow is a small team of AI agents that work together to handle conversations. This page covers the four building blocks every flow uses.

​
The four building blocks

The Start node

The single entry point. Owns the channel configuration (Widget, Gmail, Slack, etc.). One per flow.

Actors

AI agents. Each has a name, an icon, a system prompt, a model, and a set of capabilities. You can have as many as you need.

Routing edges

Connections between actors. Each edge carries an activation prompt — a short description of when this route should fire.

The canvas

The visual map where you arrange actors and draw edges. Drag, zoom, fit-to-view.

​
Start node

Every flow has exactly one Start node. It owns the channels configuration — where users reach your flow:
  • A Widget embedded on your website
  • A Gmail or Outlook inbox the flow polls for new emails
  • A Slack channel where the bot listens
  • A Teams chat
  • A phone number
A flow with no channels still works — you can test it and start conversations manually — but it won’t receive real inbound traffic until you wire at least one channel. The Start node has no AI of its own. It doesn’t reply to users; it just routes new conversations into the first actor.

​
Actors

An actor is one AI agent. Each actor is independently configured:
Triage actor config

Actor configuration panel — Triage agent with context, instructions, and definition of done filled in.

The model each actor uses is set at the workspace level. You don’t pick a model per actor — that keeps the configuration simple and lets the workspace owner upgrade models centrally.

​
Routing edges

Edges connect one actor to another. They’re how you say “from here, the conversation can go there.” Each edge has an activation prompt — a plain-language description of when this route fires:
When the user asks about price, plans, seats, discounts, free trial, or wants a quote for an Enterprise contract.
The platform reads the activation prompt during routing decisions, then auto-generates a short label (1–3 words) that appears on the edge. In this case: “Pricing”.
Edge activation panel

Edge activation panel — the long-form prompt that drives routing.

Write activation prompts as if you’re describing the situation to a teammate, not configuring software. “When the user wants a refund and has provided their order number” is much more useful than “intent = refund”.

​
How routing actually works

At each turn, the platform asks:
  1. Is the current actor still on-topic? If yes → keep talking.
  2. If no, is there a sibling actor whose activation prompt fits? If yes → hand off.
  3. Otherwise → escalate to the global router, which can pick any top-level actor or reply directly for cross-cutting questions.
You don’t write any of this routing logic. The edges + activation prompts give the platform everything it needs.

​
The canvas

The canvas is where you arrange the flow visually. It’s a free-form drag-and-drop surface.
Canvas overview

The Sales Triage flow with all four nodes and two labeled edges.

Canvas controls:

​
A simple worked example

The flow used throughout these docs: What happens when a user opens the widget and types “how much does Acme Pro cost?”:
  1. The Start node receives the message and routes it into the first actor (Triage).
  2. Triage’s activation isn’t matched on this turn — but the Pricing edge’s activation prompt clearly fits. The platform routes directly to Pricing Specialist.
  3. Pricing Specialist replies using the pricing info from global context.
You can see this trace in the Test drawer — it shows every routing decision inline.

​
Next

Context and inheritance

How information moves between actors so the user never repeats themselves.

Focus modes

How strictly an actor sticks to its job when the user changes topics.

Execution modes

Initiator vs Responder — does your flow start conversations, or wait for them?

Configure an actor

The full field-by-field guide to setting up an actor.