Build your agent

Workflows

Workflows let you script multi-step agent responses for predictable conversations — FAQ trees, lead qualification, handoff triggers — that fire before the LLM runs. They are deterministic, free of inference cost, and easy to author in a visual canvas.

Where workflows sit in the pipeline

Every visitor message goes through three filters in order:

Curated answer match — exact-question / fixed-answer pairs.
Workflow match — keyword-triggered scripts.
RAG retrieval + LLM generation — the default path for anything that didn't match above.

Activation requirements

A workflow only fires when three conditions align:

Status — must be active. Workflows start as draft for staging; disabled pauses without deleting.
Agent scope — either workspace-wide (agent_id IS NULL) or pinned to a specific agent.
Trigger match — the visitor message must match the configured keywords using one of three modes:
- any (default) — at least one keyword appears as a substring.
- all — every keyword must appear.
- exact — the trimmed, lowercased message equals the keyword exactly.

Step types

Step	Function
`message`	Send a chat bubble. Supports variable interpolation.
`question`	Pause the flow, capture the visitor's reply into a variable.
`branch`	Conditional routing based on captured values.
`tag_lead`	Append one or more tags to the conversation's lead record.
`webhook`	Fire-and-forget POST or GET to an external URL.
`escalate`	Route to a human operator inbox.

Features

Variable interpolation — reference captured values inline with {{var_name}}.
Loop guard — branches are limited to 32 jumps per turn, so a misconfigured loop can't lock the runtime.
Two editors — a visual canvas (default) for non-technical authors, a linear form view for fast edits.
Status tracking — draft / active / disabled with the same UX as agents.

Note. Bulk delete and the same selection patterns from the Agents index also work on /app/workflows.