Building agents
Testing & prompts
Two ways to find out what you have built: have the conversation, or read the instructions it was turned into.
Talking to it
The built-in tester holds a live conversation with the agent as currently saved, through your microphone, without placing a real call.
Save first
The tester runs the saved workflow, not the canvas in front of you.
Open the tester
Test in the toolbar.
Have the conversation
Speak as a caller would, including the awkward turns, an agent that handles the happy path is not yet finished.
A voice is required
What saving produces
Every save turns the graph into the prompt the agent actually runs on. It reads the nodes and their connections, expresses the structure as a diagram the model can follow, folds in each node's own configuration (prompts, variables, scripts) and writes the result, along with a preview you can read.
To see it without saving, open the Agent node's side panel and choose Update Prompt. It is worth reading before a campaign runs: the prompt is what the agent receives, and the canvas is only how you wrote it.
The layout is not decoration