Test the call path, not just the prompt

Write the caller goal and the observable outcome before making the test call. Examples include an appointment recorded, a message captured with required fields, or a completed transfer.

Test greetings, intent recognition, knowledge retrieval, required questions, tool inputs, confirmations, and the final outcome. Inspect the transcript and event stream rather than relying on how the call sounded.

Interrupt the agent, change the request, provide incomplete information, repeat an answer, stay silent, and ask an unrelated question. Confirm that the call recovers without inventing information or skipping required steps.

Use valid, invalid, duplicate, and missing inputs. Confirm that permission checks and required confirmations run. Verify the caller receives a safe fallback when a calendar, CRM, webhook, or transfer fails.

For permitted outbound follow-up, verify purpose, consent, DNC/DND status, opt-out handling, calling windows, AI disclosure, recording notice, and the audit trail. A failed gate should stop execution.

Keep goals, tools, guardrails, compliance policy, analytics, and handoff behavior in a reviewable version. If production evidence reveals a failure, update the contract, repeat the call path, and compare outcomes before republishing.