Creating a New Workflow
1
Navigate to Workflows
Go to Build → Workflows in your dashboard.
2
Click Create Workflow
Click Create Workflow to open the builder.
3
Name Your Workflow
Give it a descriptive name:
- Good: “Refund Request Handler”, “Order Status Lookup”
- Avoid: “Workflow 1”, “New Workflow”
4
Build Your Workflow
Use the visual builder to add and connect nodes.
5
Save and Publish
Save your workflow and publish when ready.
Using the Workflow Builder
The Canvas
The main area where you build your workflow:- Drag nodes to position them
- Click nodes to select and configure
- Drag from outputs to create connections
Node Palette
Available nodes you can add:- Action — Execute tools and operations
- Decision — Branch based on conditions
- AI Decision — LLM-based branching
- AI Extract — Extract data with AI
- Respond — Send messages
Properties Panel
Configure the selected node:Adding Nodes
Action Nodes
Execute a tool or operation:1
Add Action Node
Drag an Action node to the canvas.
2
Select Tool
Choose which tool to execute (Knowledge Search, API call, etc.).
3
Configure Inputs
Map data from previous nodes or the conversation to the tool’s inputs.
4
Name Outputs
Define output variable names for use in later nodes.
Decision Nodes
Branch based on conditions:1
Add Decision Node
Drag a Decision node to the canvas.
2
Define Condition
Write the condition to evaluate (e.g.,
order.status == 'delivered').3
Connect Outputs
Connect the Yes and No outputs to different paths.
AI Decision Nodes
Let the LLM make a judgment call:1
Add AI Decision Node
Drag an AI Decision node to the canvas.
2
Write Prompt
Describe what the AI should evaluate (e.g., “Is the customer satisfied with the resolution?”).
3
Define Options
Create the possible outcomes (e.g., “Satisfied”, “Not satisfied”, “Unclear”).
4
Connect Outputs
Connect each option to its respective path.
AI Extract Nodes
Extract structured data from conversation:1
Add AI Extract Node
Drag an AI Extract node to the canvas.
2
Define What to Extract
Describe the data to extract (e.g., “order number”, “customer email”).
3
Configure Output Schema
Define the structure of extracted data.
Respond Nodes
Send a message to the customer:1
Add Respond Node
Drag a Respond node to the canvas.
2
Write Response
Write the message template. Use variables from previous nodes:
Connecting Nodes
Creating Connections
- Click and drag from a node’s output port
- Drop on another node’s input port
- The connection is created
Connection Rules
- Every node (except Start) needs at least one input connection
- Decision nodes have multiple output ports (one per branch)
- Action nodes have Success and Failure outputs
Configuring Inputs
Variable Mapping
Map data between nodes using the properties panel: Available sources:| Source | Description | Example |
|---|---|---|
| Conversation | Data from the message | conversation.latest_message |
| Previous node | Output from earlier nodes | extract_order.order_id |
| Static value | Hardcoded value | "pending" |
Testing Workflows
In the Builder
- Click Test in the builder
- Enter test inputs
- Run the workflow
- See which path was taken and outputs at each step
In Playground
- Assign the workflow to an agent
- Test in Test → Playground
- Conversation-level testing with full context
Publishing Workflows
Workflows start as drafts. To make available to agents:- Click Publish in the builder
- Confirm the publication
- Workflow is now available for agent assignment
You can continue editing a published workflow. Changes create a new draft that must be published to take effect.
Best Practices
Design
- Start simple — Build the happy path first, then add branches
- Name clearly — Use descriptive names for nodes and variables
- Group logically — Organize related nodes spatially
Testing
- Test each branch — Ensure all paths work correctly
- Test edge cases — What happens with missing data?
- Test failures — How does the workflow handle errors?
Maintenance
- Document decisions — Add notes explaining complex logic
- Version carefully — Test changes before publishing
- Monitor execution — Check run history for issues