OpenAI voice agent exposes pipeline gap

This reads like a classic demo-to-production wake-up call: voice agents fail at the control-loop layer, not the generation layer. OpenAI’s own voice docs basically back that up by treating turn detection and interruption handling as first-class system decisions.

• –Real-time voice needs barge-in handling, turn detection, response cancellation, and audio truncation; prompts alone cannot cover those behaviors.
• –The same issue across multiple vendors suggests the bottleneck is system design and latency, not one model family.
• –OpenAI’s docs separate manual and automatic turn detection and describe VAD-driven interruption handling in Realtime.
• –Builders should test with live interruptions, pauses, and mid-sentence topic shifts, not just clean scripted turns.
• –If a voice agent sounds good in a playground but fails on calls, that usually means the stack is missing state, not intelligence.