Steps & nodes
The node reference: dialogue, knowledge, webhook, notify, transfer, logic split, extract, function, keypad, wait, end — plus transition conditions, variable extraction, global nodes, and per-step tools.
View as MarkdownEvery step in a workflow is a node of one of these types, and every line between two steps carries a transition condition that decides when the conversation moves along it. Some steps talk; some just do a job and move on.
Node types
| Node | What it does |
|---|---|
| Dialogue | The AI speaks for this step — from a per-step prompt, or exact static wording when the words are safety- or brand-critical. Can extract typed variables (below). |
| Knowledge | Pins reference material to the step — the menu, the price list, the policy. Only in play on this step, so long documents don't bleed elsewhere. |
| Webhook | Calls any API: method, URL, headers, body with {{variable}} interpolation; response values mapped back into variables via JSON paths; conditional routes on the result (paid → this way, failed → that way). POST /crm/ingest with a Connections key writes a row + deal into the CRM. |
| Send a notification | Email, SMS, WhatsApp, or a help-desk ticket — see Notifications. |
| Transfer | Hands the call to a human number, with context. On chat, hands to the help desk. |
| Transfer to workflow | Jumps to another workflow — compose big flows from small ones ("refill requests" as its own flow). |
| Logic split | Branches the moment it is reached, without speaking. Compares the variables collected so far and sends the conversation down the matching path. Always give it an else way out — the editor refuses to publish one without. |
| Extract a variable | Pulls named values out of what was just said and moves on without speaking. Types: text, number, yes/no, or a fixed list. |
| Function | Runs the tools attached to the step, for when calling out is the step — book the slot, look the order up. Nothing to say; put a dialogue step after it to read the result back. |
| Press a digit | Asks for keypresses and stores them. On a call the phone line gathers the tones; set how many digits, how long to wait, and what key ends it. Calls only. |
| Subagent | Dialogue where the AI decides whether and when to reach for a tool, from what the customer says. Use a function step instead when the tool must run every time. |
| Call an MCP tool | Calls one tool on an MCP server of your own. Takes the same request headers and response mapping a webhook step does; a {{SECRET.name}} header resolves at the moment of the call and is never logged. |
| Wait | Pauses the flow. |
| End | Closes cleanly with a final line. |
Transition conditions
Each line leaving a step says how it is taken. Mixing them is normal: check what you can check, and leave the rest to judgement.
| Condition | When it is taken |
|---|---|
| When the model judges… | You write the condition as a sentence — the caller said they want to book. Judged together with the other sentence conditions, so the model picks one best match or none. |
| When the numbers say so | Rows of variable / operator / value, combined with ALL or ANY. No model involved. |
| Always, once they reply | Taken as soon as the caller says anything. Nothing else on that step is looked at. |
| Otherwise | The fallback, taken when nothing else matched. |
| As soon as the agent stops speaking | Moves on without waiting for a reply — for a line that needs no answer. |
The order they are judged in
This order is fixed, and it matters: anything that can be decided for certain is decided before anything is left to the model.
- 1Always — if the step has one, it wins outright.
- 2Equations — top to bottom, the first one that holds.
- 3Sentence conditions — all judged together, alongside any global step.
- 4Otherwise — the fallback.
- 5Nothing matched: the conversation stays where it is and keeps talking.
Operators
| Operator | Compares | Behaviour |
|---|---|---|
| is / is not | Text | Ignores case, and never reads text as a number — so 18 and 18.0 are different values. |
| > ≥ < ≤ | Numbers | If either side is empty or is not a number the row is simply false. It never errors. |
| contains / does not contain | Text | Whether one value appears inside the other, ignoring case. |
| is set / is not set | Anything | Whether the variable has a value at all. An empty answer still counts as set. |
Variables
A dialogue step or an extract step can declare typed variables — {name} (text), {party_size} (number), {is_emergency} (yes/no), or a fixed list where an answer outside the choices is refused rather than stored — and the AI extracts them from the conversation. Extracted variables flow into every later step: webhook bodies, notification wording, the deal card. Built-ins are always available: the caller's number, the time, the last user message.
Global nodes
A global node listens across the whole flow, not one position in the graph. The canonical one: "caller asks for a person" → transfer. Wherever the conversation is, the escape works. Give every customer-facing flow at least that one.
Two settings make one behave: go back afterwards returns the conversation to wherever it was once your go-back condition reads true, and stay quiet for N steps takes the step out of the running for a while after it fires, so a customer who keeps phrasing things the same way is not bounced straight back into it.
Per-step model
A step can name its own model. Leave it empty and the step uses the agent's. Naming a cheaper one on a routing step is how the cheap parts stay cheap while the step that matters gets something stronger.
Worked examples
When a step words something oddly, or keeps taking the wrong way out, give it examples: a snippet of transcript, and what should have happened. Response examples steer what it says; transition examples name the step it should have reached.
Per-step tools
Tick a tool on a step and only that step can use it — booking on the booking step, the POS push on the order step, Stripe lookup on the billing step. Tight scope is what keeps tools from firing at the wrong moment.
For the shut-off instructions, the allergy disclaimer, the legal line — set the dialogue node to exact wording. The AI reads it verbatim, every time.
Every node type and every transition condition can be set through create_workflow and update_workflow_graph — edges take condition_kind, equation_match and equations. See the MCP tools reference.