Use this checklist for the exact draft you intend to publish. Record the agent ID, draft version, purpose of the change, and the run IDs used for verification. If the draft changes after testing, repeat the checks affected by that change.
Definition and change scope#
- The implementation assistant fetched the current workflow code before editing and preserved unrelated nodes, edges, prompts, variables, tools, and documents.
- Every new or changed node uses the current node-type schema and has a clear name, focused prompt, supported fields, and reachable edges.
- Edge labels are short and edge conditions are complete, observable statements about the live conversation.
- save_workflow returned saved: true for the intended agent ID and produced the draft version being reviewed.
- The published workflow remained unchanged while the draft was under test.
Models, voice, and conversation#
- The model configuration is valid for the deployment: realtime speech-to-speech plus analysis LLM, Menace Voice-managed inference, or a complete BYOK stack.
- The opening identifies the agent appropriately, establishes the purpose, and does not make claims the workflow cannot support.
- Required information, confirmation points, success criteria, and closing behavior are explicit in the owning nodes.
- Normal speech, silence, interruption, correction, ambiguity, out-of-scope requests, and the configured escalation path were exercised.
- The transcript, audio, node transitions, and final response were reviewed together rather than inferred from configuration.
Data, tools, and external systems#
- Initial-context producers and prompt references use the same variable names, and missing values have an intentional fallback.
- Gathered-context fields have stable names, correct types, evidence-based extraction instructions, and downstream consumers that handle empty values.
- Documents are processed, current, and attached only to nodes where they are relevant; Chunked Search has a working embedding configuration.
- Tools are attached only to the nodes that need them, credentials are referenced from secure storage, and no secret appears in prompts or context.
- Success, not-found, validation, authorization, timeout, and duplicate-action behavior were tested for every changed external tool.
- Webhook payloads were received and validated by the intended backend, with the expected context, disposition, and artifact URLs.
Delivery path#
- Website delivery: the selected widget mode loads, obtains microphone permission when required, starts a run, and behaves at desktop and mobile widths.
- Inbound telephony: the active number is assigned to the correct agent and a real inbound test reaches the expected published workflow.
- Outbound telephony: the configuration is active and ready, the caller ID is valid, and a real test connects through the selected provider.
- API trigger: the correct trigger UUID or Agent UUID route is used, X-API-Key scope is correct, and the returned run ID is followed to a terminal result.
- Campaign: a small contact batch proves CSV mapping, time windows, concurrency, retry rules, circuit breaker behavior, and result reporting before scale-up.
Operations and release#
- Run status, transcript, recording, initial context, gathered context, usage, cost, tool activity, and errors are visible to the operator who owns the launch.
- QA output and call dispositions match transcript evidence and are not treated as a substitute for reviewing the first real calls.
- The owner knows when to pause traffic, what failure threshold matters, and how callers reach a human or receive a clear next step.
- The reviewed draft is published once, then a production-path smoke call confirms the published version rather than the draft route.
- The previous published version and the change record remain available for audit and recovery.
Prove the boundary
Structural validation, a saved draft, a Web Call, a telephony initiation response, and a completed production call are different evidence layers. Report exactly which layers passed and do not describe an unobserved layer as complete.
Return to Build agents that can be trusted when starting the next bounded workflow change.