A realtime agent says an interrupted action already happened

A customer interrupts a voice assistant while it is submitting a refund. The assistant says the refund has already been submitted, although the HTTP request has not yet reached the service. An OpenAI Agents SDK issue reports this behavior with asynchronous tool calls during a realtime interruption. The reporter's repeated local tests show a reproducible symptom in that setup; they are not a general reliability rate for every deployment.

Make action state explicit

Represent each consequential tool call as requested, running, confirmed, failed or unknown, with an operation ID. Only a confirmation from the business system can move it to confirmed. A spoken intention or a tool invocation scheduled in the runtime is not completion. The Realtime Agents guide covers session and tool behavior; the application must still define what its customer-facing status words mean.

Handle interruption as a reconciliation event

When speech is interrupted, continue tracking the outstanding operation independently of the response being cut short. Before answering “did you do it?”, query the operation record or target API. If the state is unknown, say so and offer to check rather than retrying blindly. For refunds, bookings and account changes, use idempotency keys where the target service supports them so a retry does not create a second side effect.

Test the dangerous timing window

Delay the HTTP call deliberately and interrupt at several points: before dispatch, during transit and after confirmation but before the response is spoken. Verify the customer's answer and the underlying business record at each point. Turning off asynchronous calls may change latency; that tradeoff should be measured with a real conversation, not assumed to be free.

Adovo’s view

Trust is lost in the gap between “I will do it” and “it is done.” An agent should speak from verified state, especially when the customer's money or account is involved.