Every step in a flow is a node, and every way out of a node is an edge. This page covers what all nodes have in common. Each node type has its own page for its specific fields.
Common node fields#
| Field | Type | Default | Description |
|---|---|---|---|
id |
string | Unique within the flow. The dashboard generates node_... ids |
|
type |
string | The node type | |
name |
string | "" |
Shown on the card and used in search, logs and test output |
display_position |
{x, y} |
{x: 0, y: 0} |
Canvas position |
global_node_setting |
object | Makes the node a global node | |
instruction |
{type: "prompt" or "static_text", text} |
On most types. See below |
Instructions#
Nodes that speak have an instruction:
- Prompt (
type: "prompt"): text for the LLM, added to the system prompt under "Current Task". The LLM writes the words. - Static Sentence (
type: "static_text"): exact words. A conversation node speaks them verbatim on entry, then tells the LLM it has just said them.
Both render {{variables}}. Instructions exist on conversation, subagent, press digit, end, transfer, function, code, SMS, agent transfer, MCP and transfer-agent nodes. On a function node the field is labelled Message while running (optional).
Edges#
An edge is { id, destination_node_id, transition_condition }. destination_node_id is null while the edge is not connected; an unconnected edge is ignored at runtime.
| Edge kind | Field | Where | Taken when |
|---|---|---|---|
| Normal | edges[] |
Most types | Its condition holds |
| Always | always_edge |
Conversation, subagent | After any caller reply, before other conditions |
| Skip response | skip_response_edge |
Conversation | Right after the agent finishes speaking, without waiting for the caller |
| Else | else_edge |
Conversation, subagent, function, code, MCP, component; required on logic split | No other edge matched |
| Failure | edge |
Call transfer (Transfer failed), agent transfer (Swap failed) | The transfer or swap failed |
| Success and failure | success_edge, failure_edge |
SMS (SMS sent, SMS failed) | The SMS was or was not sent |
On a node card, the Transition section lists every edge. The + button (Add transition) offers Prompt, Equation, Else and Always/Skip where the node type allows them. In Node settings, conversation nodes also have Skip response, Always edge and Else edge toggles.
Evaluation order#
When the engine decides whether to leave a node, it checks in this order:
- Always edge.
- Equation transitions, top to bottom. The first that holds wins.
- Prompt transitions, Both (equation plus prompt) transitions whose equations hold, global node conditions and go-back conditions, decided together by the LLM.
- Else edge.
- Otherwise stay in the node.
When the checks run depends on the node:
| Node | When transitions are checked |
|---|---|
| Conversation, subagent without an else edge | Steps 1 and 2 before the LLM replies to each caller turn. Step 3 happens inside the reply call: the LLM either answers or calls a transition_to_<node_id> tool |
| Conversation, subagent with an else edge | All steps before the reply, with one extra LLM call for step 3, so that "else" can apply |
| Every other node | All steps right after the node finishes its work. Step 3 uses a separate LLM call |
| Skip response edge | After the agent finishes speaking in the node |
The engine follows automatic transitions for up to 25 hops in a row; after that it stays put and logs an error, which protects against loops.
Per-node settings#
Open Node settings at the bottom of the node's settings to override agent or flow settings while the node is active. Empty fields mean "same as the agent" or "same as the flow".
| UI label | Field | Range |
|---|---|---|
| Model | model_choice: {type: "cascading", model} |
Any listed model, or Same as the flow |
| Voice speed | voice_speed |
0.5 to 2 |
| Responsiveness | responsiveness |
0 to 1 |
| Interruption | interruption_sensitivity |
0 to 1 |
| Transcription | stt_mode |
Optimise for speed (fast) or Optimise for accuracy (accurate) |
| Boosted keywords | boosted_keywords |
Comma separated |
| Reminder after (s) | reminder_trigger_ms |
1 s or more |
| Reminders, at most | reminder_max_count |
0 or more |
| Keypad input interrupts the agent | allow_dtmf_interruption |
Yes, No, or same as the agent |
| Retrieval instruction | kb_instruction |
Up to 500 characters |
| Knowledge bases | knowledge_base_ids |
Used instead of the flow's knowledge bases |
| Example replies, Transition examples | finetune_conversation_examples, finetune_transition_examples |
See Finetune examples |
The model override also accepts high_priority and ultrafast through the API. When the next node has no override, the agent's own value is restored.
Start node#
The start node is start_node_id. On the canvas the green Begin tag points at it, and the node's header shows Start in Node settings. Change it with Set as start node in Node settings, or drag a line from Begin to another node. Deleting the start node makes the next node in the list the start.