A caller repeating a request may be checking progress rather than authorising another action. Test whether the agent preserves one business operation instead of creating a new write every time it hears similar words.
Repetition is not always duplicate intent
Repeated-request testing checks how the agent distinguishes clarification, status checks, restated intent, and genuinely new actions. The wording alone may be identical while the operational meaning differs.
In a proposed state model, keep the current requested operation, confirmation revision, and execution status separate from the transcript. When a caller repeats a pending request, the agent can inspect or explain that operation rather than submit it again.
Do not rely on text similarity as the sole duplicate rule. Two appointments can have similar descriptions but different dates or recipients. Conversely, different wording can express the same operation after a response delay.
Use a fictional delayed booking
The caller confirms a Wednesday visit. While the booking service is slow, they ask, “Can you book Wednesday?” The system should recognise the pending operation and avoid a second create request. Once the result arrives, it can confirm the existing appointment.
Now test “Book another visit for Friday as well.” That is potentially a new request, not a duplicate. The agent should collect the necessary details and confirmation under the business’s policy rather than suppress it because a booking already exists in the session.
| Repeated-request variant | Expected handling |
|---|---|
| Same request while pending | Explain or check the existing operation |
| Same request after completion | Read back the existing result |
| Restated request with corrected date | Resolve the changed intent explicitly |
| Explicit additional appointment | Treat as a new proposal |
| Repeated cancellation after success | Do not create another reversal effect |
Inspect operation identifiers and side-effect counts. A single final confirmation can conceal multiple backend writes if the tool service does not enforce duplicate protection.
Test speech and transport duplication separately
A caller can intentionally repeat themselves, and a network or event pipeline can also deliver a duplicate message. Both can lead to repeated execution, but they have different causes. Include fixtures for duplicated events as well as conversational repetition.
Overaggressive suppression can block legitimate additional work. Overly permissive execution can create duplicates. Use explicit operation identity and targeted clarification when intent changes, rather than a broad rule that every repeated phrase is ignored.
This test design is proposed. The voice platform case study describes related tool work. Read booking idempotency for the backend guarantee that complements conversational handling.
Repeated-request FAQ
Should the agent always ask whether this is a duplicate?
No. If the current operation clearly matches, explain its status. Clarify when consequential details differ.
Can the model alone prevent duplicate writes?
No. The execution service needs an appropriate operation-identity and concurrency mechanism.
What if the caller says the previous booking failed?
Check the authoritative operation or appointment state before creating a replacement.