Hedge side-effect-free LLM routing calls to cut voice tail latency

In a cascaded voice agent, a rare slow routing response from a hosted LLM can leave the caller in multi-second silence, even when the request is small and fully cached. If routing only selects an action and business handlers run after the result is committed, the call has no side effects, so it is safe to hedge it. When no answer has arrived by roughly the measured p95, send one identical duplicate, use whichever succeeds first, and cancel the other. Cancel both on caller interruption. The extra cost lands only on the slow tail. Limits: the hedge must never duplicate a handler or tool execution. A filler "okay" masks the delay, but it does not remove it.

Applies when: a speech pipeline waits on a tool-choosing LLM call before it can speak. Keywords: LLM tail latency hedged request duplicate cancel voice agent OpenAI Responses routing silence