Writing the call workflow

Most people write a voice agent's script as one long instruction and then wonder why the call wanders. The Workflow section fixes that by making the call an ordered set of stages, each with a job to do, a budget to do it in, and a defined way out. It is the difference between a prompt and a script.

Set the call goal first

Open the agent version, expand Agent configurationWorkflow, and fill the two fields at the top before you add a single stage.

Call goal is the overall objective — the one thing the call exists to achieve. Approved next steps is the list of calls-to-action the agent is allowed to offer, one per line. The placeholder shows the shape:

Schedule a meeting
Request a callback
End politely

Anything not on that list is not on the table. That is the point of the field: it is a closed set, not a suggestion.

Add the stages

Use Remove stage to delete one.

The ten stages

They run in this order, and the names describe what each is for:

Stage	What it is for
Introduction and disclosure	Greeting, who is calling, and the AI disclosure
Permission to continue	Asking whether now is a good time
Discovery	Understanding the situation
Qualification	Establishing whether this is a fit
Value alignment	Connecting what they need to what you offer
Objection handling	Answering the pushback
Call to action	Offering one of the approved next steps
Booking or transfer	Actually booking, or handing to a person
Confirmation	Reading back what was agreed
Wrap up	Closing the call cleanly

You do not need all ten. A callback-confirmation call might be four. A cold discovery call might be eight.

Settings reference

Budgets are the part people get wrong

A stage's Maximum turns and Maximum seconds are a budget, not a target. Four turns is enough to ask a question, hear an answer, ask a follow-up and acknowledge it. If a discovery stage needs twelve turns, the objective is probably two objectives wearing one coat — split it.

The Fallback behavior matters more than it looks. It is what the agent does when the budget runs out with a required slot still empty. The default — Move on politely without inventing information — is the right instinct: the failure mode you are guarding against is an agent that invents the answer rather than admit it did not get one.

When stage supervision actually applies

This is the part that surprises people. Writing stages always shapes the composed prompt, but deterministic stage supervision — the runtime actually enforcing the turn and time budgets, the listen-first timing, the interruption handling, the bounded memory and the silence policy — only happens when Workflow-supervised voice runtime is switched on in SettingsVoice & Telephony.

Troubleshooting

Symptom	Cause	Fix
A stage you added has the wrong name	Adding a stage auto-selects the first unused stage name	Change the Stage select on that row
The agent ignores a stage's turn budget	Workflow-supervised voice runtime is off for the workspace	Turn it on in Settings → Voice & Telephony; it applies to new calls
The agent invents an answer when it runs out of turns	Fallback behavior was rewritten to push harder	Restore a fallback that moves on without inventing information
Validation fails against Workflow when you publish	A stage is missing a required field, or a budget is outside its range	The error names the field — fix it and validate again
The agent offers something you never approved	It is not in Approved next steps, but the persona or objective implies it	Keep the offer in the approved list and out of free-text fields

Related

• Create your first voice agent
• Test an agent before it calls anyone
• Voice & Telephony settings
• Build a knowledge base for your agents