A Shopbell voice agent is a graph. Each node is a moment in the conversation. Each path between nodes is a possible next step. The caller speaks. The current node interprets. The agent picks the next node based on rules. That is the entire model. This guide covers the four node types you'll use most and how they connect.
The four core node types
Start Call: The entry point. Every workflow has exactly one. The Start Call node fires the moment a call connects (Twilio answers, the Web Call starts). It speaks the greeting and routes to the first Agent node.
Agent: The brain. An Agent node runs the LLM. It listens to the caller, generates a response, speaks the response, and routes to the next node when the caller's intent changes. A workflow usually has multiple Agent nodes — one for greeting/discovery, one for objections, one for closing — each with its own system prompt.
Transfer: Hands the call off. A Transfer node forwards the live call to a phone number — your cell, a front desk, an outside line. Used for escalation, emergencies, and 'I want to talk to a human' requests.
End Call: Closes the conversation. The End Call node speaks the closing line and hangs up. Anything in this node's system prompt becomes the closing words — 'Thank you for calling, have a great day' or 'You'll receive a confirmation text shortly' or 'I've notified the on-call plumber, expect a callback within 15 minutes'.
How nodes connect with paths
Each node has one or more output paths. Click a node in the builder, then drag from the small circle on its right edge to another node. That creates a path. Multiple paths from one node mean the LLM can pick which one to follow based on what the caller said.
Example: an Agent node for greeting has two paths. Path A goes to a 'Book Appointment' Agent node if the caller wants to schedule. Path B goes to the Transfer node if the caller asks for a human. The Agent node's system prompt tells the LLM which path to pick. Paths can have labels like 'Caller wants to book' and 'Caller asks for human' so the LLM knows the criteria.
The Global node
Every workflow has one optional Global node — a special Agent node whose system prompt is appended to every other Agent node's prompt. Use it for rules that should always apply, no matter what node the caller is in. Examples: 'Always disclose you are an AI assistant when asked.' 'Never quote a final price in-call — say the price is in the written contract.' 'If the caller says stop, transfer immediately.' See the Global Node guide for patterns.
What goes in the Start Call node
The Start Call node has two settings. The greeting — the first thing the agent says. The first Agent node to route to. Edit the greeting to match your business. The greeting is the most important three seconds of the entire call — it sets the tone. Keep it short, friendly, and clear. Example: 'Thanks for calling Bloom Floral. How can I help you today?'
Frequently Asked Questions
Can a node route back to a previous node?
Yes. Paths can loop. A common pattern: Agent node for 'objection handling' routes back to the main Agent node once the objection is resolved. The LLM knows the conversation history, so it does not repeat the greeting.
What happens if no path matches?
The current Agent node keeps talking. The Agent node continues the conversation using its system prompt. Only when an explicit 'transfer to node X' rule fires does the conversation move.
Can I have multiple Start Call nodes?
No. Every workflow has exactly one Start Call. If you need different greetings for different times of day, use the time field in the Agent node's system prompt (e.g. 'If it is after 5pm, say 'thanks for calling, we are closed but I can take a message').