Flow builder

Visual canvas for the agent-level conversation graph.

The Flow tab shows the top-level graph. In intent-first designs, each intent owns its own sub-flow; the main flow often contains a single intent entry node. Use the canvas to wire high-level routing or legacy monolithic flows.

Canvas basics

  • Drag nodes from the library onto the dotted grid.
  • Connect handles to create edges; click an edge to delete or label it.
  • Select a node to edit properties in the inspector (right panel).
  • Every node has a Name field — used in the execution trace (mandatory for clarity).
  • Changes autosave after a short debounce; use Save for an immediate write.

Node types

NodePurpose
Intent / StartEntry point for a conversation path.
MessageBot reply; per-channel copy and rich UI (buttons, cards, carousel, list).
EntityPrompt user and capture a slot value.
DecisionBranch on edge conditions (yes/no or expressions).
Service / ToolHTTP call with mapped outputs.
PythonSandboxed script for custom logic.
SetAssign a session variable.
HandoffTransfer to a human agent.
SubflowInvoke another flow by reference.
KnowledgeRetrieve chunks and optionally synthesize an answer.
instruction_agentEmbedded AgenticLoop (LLM + tools) — see Runtime engine.
EndTerminate the session.

Node repair

Inspector or context menu can trigger scoped LLM repair for a single node (POST /v1/agents/{id}/flow/repair-node). Requires LLM extraction configured.

Dragging a node only updates its position. Connections are preserved when you save — if edges disappear after reload, use Versions to restore a known-good draft.