An inbound agent deployment is ready when a call reaches the intended implementation and produces recoverable business outcomes. A container starting or a worker registering does not establish the whole phone workflow.
LiveKit's custom deployment guide describes agent-server registration, resource-aware availability, and graceful rollout. Verify those behaviors against your installed version and hosting platform. The checklist below is a proposed acceptance structure, not a claim about an already-tested deployment.
Verify the route and identity first
Record the phone entry point, inbound configuration, selected agent, environment, business configuration, and remote destinations. Use independent staging credentials and controlled records. A development agent should not accidentally accept real client traffic.
Check secrets and outbound dependencies without printing credentials into logs. Confirm that the deployment can reach the required model and business services and that rejected access produces an honest fallback rather than a false booking confirmation.
Rehearse a fictional service-office launch
The agent handles routine intake and a narrow booking flow. Acceptance calls include a known customer, two possible properties, no available visit, an unanswered transfer, a failed API write, and a caller who disconnects during submission.
Reviewers inspect the destination records and unresolved queue. The launch record identifies which cases were exercised against real controlled integrations and which were simulated.
| Deployment check | Evidence required |
|---|---|
| Inbound route | Intended room and agent context |
| Business scope | Correct account and destination |
| Resource pressure | New work handled under approved policy |
| Rolling release | Existing sessions treated as planned |
| Abrupt process loss | Durable request and reconciliation |
| Alerts | Named staffed recipient and fallback |
Treat rollout and crash recovery differently
Graceful draining can preserve active work during a planned update when the hosting grace period supports it. An abrupt process loss can terminate in-memory work. Persist important action evidence and define the customer recovery process for that case.
A conservative deployment may need spare capacity and a staffed review window. Aggressive scaling or rapid termination can reduce cost while increasing interruption risk. Measure your actual workload rather than copying a generic concurrency figure.
Keep rollback instructions for the complete behavior package, including action mappings and configuration. Review voice launch checks for operational ownership around deployment.
Questions about inbound deployment
Is worker health enough to launch?
No. Test the real phone path, business actions, failure queues, and staff handoff.
Should readiness depend only on CPU?
Choose verified resource and dependency signals appropriate to the workload and platform, then test their behavior under pressure.
What belongs in the launch pack?
Versioned configuration, environment and account mappings, acceptance evidence, known limits, monitoring owners, rollback, and the recovery procedure for interrupted calls and uncertain actions.