Agent Builder Guide

Master the visual agent builder to create powerful conversation automations — or let the built-in AI assistant build them for you in plain language.

AI Agent Assistant

New

Describe what you want — the assistant builds it for you

The Agent Assistant is a Copilot-style chatbot built into every agent page. You describe your intent in natural language and it does the work: it reads the current agent, creates or edits nodes, draws the right edges, installs tools from the catalog, and writes the AI prompts for you. Every change is snapshotted, so a single click rolls back to any previous state.

You don't need to drag a single node to get started. Open the assistant, type one sentence, and you'll have a working draft of your agent in seconds — then refine it in plain language or polish manually on the canvas.

What it can do

▸Build a complete agent from a description — nodes, edges, prompts, and conditions all included
▸Edit any existing node (label, position, system prompt, config field, branching rules) with surgical precision
▸Install tools from the catalog (Google Calendar, CRM, HTTP endpoints, etc.) and wire them into your AI nodes
▸Drop in a Tool node that calls a tool directly, without an AI agent: each parameter can be static, taken from the context, or filled by a single structured AI call you can prompt
▸Auto-snapshot the agent before every change — one click in the chat to restore any previous state
▸Streams its reasoning token-by-token and shows every tool call it makes, so you always know what it just did
▸Conversations are persisted: close the tab, come back days later, and pick up exactly where you left off

Example prompts

"Build an agent that greets the user, asks for their email, and saves it to the contact."
"Add an AI node after the welcome message that answers from the knowledge base."
"Connect a Google Calendar tool to the AI node so it can book meetings."
"Change the system prompt of the support node to be more concise and ask for an order number first."
"Add a Tool node that calls 'schedule_appointment': pull the name and phone from the contact, let AI infer the date and reason from the conversation."

How to use it

  1. 1Open any agent in the Agent Builder
  2. 2Click the Assistant icon at the top-right of the canvas
  3. 3Type a request — e.g. "add a payment confirmation message after the order node"
  4. 4Watch the canvas update live as nodes appear and edges are drawn
  5. 5If something is off, hit ↩ Restore on any past message to roll the whole agent back to that point

Node Types

Start

Entry point of the agent. Triggered when a message is received.

Message

Send a text, image, or document message to the user.

Condition

Branch the agent based on message content, variables, or contact data. Supports 9 condition types.

AI Response

Generate a response using GPT-4o, Gemini, or Claude. Optionally include conversation history.

RAG Query

Search the knowledge base for relevant content and use it to answer the user query.

Tool

landing.docs.agentBuilder.nodeTypes.tool.desc

Collect Input

Wait for user input and store it in a variable. Supports text, number, email, and phone validation.

Set Variable

Create or update a variable. Use expressions to compute values from other variables.

Delay

Pause the agent for a specified duration. Useful for typing indicators or scheduled messages.

Recurring Reminder

Repeat a message or action on a fixed interval or cron schedule until the user replies, a deadline passes, or a max count is reached. Non-blocking — the agent continues immediately.

Loop

Iterate over a list inside the conversation (internal) or spawn one separate agent run per matched contact or conversation (spawn).

landing.docs.agentBuilder.nodeTypes.startFlow.name

landing.docs.agentBuilder.nodeTypes.startFlow.desc

Transfer Agent

Hand off the conversation to a human agent. Includes context and conversation history.

Note

Add a documentation note on the canvas. Does not affect agent execution.

End

Terminates the agent. Optionally mark the conversation as resolved.

Condition Types

The Condition node supports 9 comparison types for flexible branching logic:

Contains textExact matchStarts withEnds withRegex patternNumber greater thanNumber less thanIs emptyDefault (fallback)

Best Practices

  • Start simple — build a basic welcome agent first, then add complexity
  • Use the AI Agent Assistant to scaffold your first draft, then refine on the canvas
  • Use the Condition node to handle common intents (billing, support, sales)
  • Always include a default/fallback path for unrecognized messages
  • Use RAG Query before AI Response to ground answers in your data
  • Add a Transfer Agent node for cases where AI can't resolve the query
  • Test your agent with the Chatbot Playground before deploying
  • Use Note nodes to document complex branching logic
  • Keep agents modular — use separate agents for different purposes

Edge Types (Transitions)

Edges connect nodes together. The Agent Builder supports 6 edge types for different connection styles:

Default
Success
Failure
Condition True
Condition False
Timeout