Launch an Agent: Production Checklist
Take a GetVoiceBot agent from a defined business outcome through testing, publishing, channel activation, and a controlled production rollout.
A successful launch proves more than whether an Agent can answer a test question. The Agent must
understand its job, complete approved actions, handle uncertainty, use the correct communication
channel, report useful outcomes, and give your team enough evidence to support it.
1. Define the outcome
Write down one primary workflow before configuring the Agent:
- who will contact or receive contact from the Agent
- what successful completion means
- what information the Agent needs
- which actions it may take
- when it must escalate or stop
- what your team will measure after launch
Start with one measurable workflow. Add adjacent responsibilities only after the primary path is
reliable.
2. Configure the Agent
In Edit Agent, review the complete prepared configuration:
- instructions, greeting, model, and reasoning settings
- language, transcription, and voice behavior
- functions and connected apps
- links and approved knowledge
- reporting fields
- transfer, notification, and escalation behavior
Keep credentials out of instructions. Connect them through the supported app, skill, or managed
credential flow.
3. Connect the delivery channel
Publishing and channel setup are separate. Choose the routes the workflow actually needs:
- assign or bring a phone number for calling and SMS
- complete SMS registration before application-to-person messaging
- configure SIP when calls will use your telephony infrastructure
- publish a profile or website embed for web voice and chat
- connect email or supported messaging channels where enabled
See Phone numbers and channel routing.
4. Test the prepared configuration
Test through the same channel customers will use. Cover at least:
- the ideal successful path
- missing or ambiguous information
- an unsupported request
- escalation or human handoff
- each function, app, transfer, message, or other external side effect
- a reporting outcome for success, failure, and uncertainty
Verify every external action in its destination system. An Agent saying it completed an action is
not proof that the action occurred.
5. Prepare measurement and recovery
- Add only the reporting fields needed to evaluate the workflow.
- Confirm the metric population and denominator.
- Save a named version before a major release.
- Decide which result would trigger rollback or escalation.
- Identify the person responsible for reviewing early conversations.
See Measure and improve your Agent.
6. Publish and verify live behavior
Publish the tested configuration, then reopen the live Agent and confirm the intended settings.
Place a real test call or message through the production route and inspect its log and complete
conversation.
Publishing does not automatically assign a number, complete SMS registration, change SIP routing,
or notify customers.
7. Roll out deliberately
Start with a controlled audience, location, queue, or traffic source. During the first production
period:
- review representative successful and unsuccessful conversations
- watch transfers, external actions, latency, and channel failures
- compare reporting values with the underlying conversations
- fix changes in the prepared configuration, test again, and republish
Do not change several major behaviors at once when you need to understand which change improved or
degraded performance.
Launch acceptance criteria
The Agent is ready when:
- its live configuration matches the approved prepared version
- every intended channel reaches the correct Agent
- the primary workflow and escalation paths pass representative tests
- external actions are confirmed in their destination systems
- logs and conversations contain enough evidence to investigate a problem
- reporting fields and metrics produce understandable results
- a recoverable prior version exists for material changes
For configuration details, see Set up a GetVoiceBot agent and
Publish and share your Agent.