API: interruption sensitivity on the agent

Agent config gains turn_handling.interruption_sensitivity, controlling how easily a caller takes the floor from a speaking agent and how patiently the agent waits before treating a pause as the end of a turn. The value is one of low, medium, high, or null.

null follows the platform default and behaves identically to medium. Nothing you already send changes — an agent with no value set behaves exactly as before.

  • low makes the agent hard to interrupt: background noise, hold music, and one- or two-word sounds no longer stop it mid-sentence (it needs ~1.5s of speech and at least 3 words), and it waits ~0.4s longer before taking a pause as the end of a turn. Use it for noisy lines and for callers who pause mid-thought.
  • high is the opposite: the agent yields after ~0.4s of caller speech and takes the turn sooner.

Neither setting changes ordinary turn-taking — a one-word answer still replies normally when the agent is silent; the word threshold only applies to speech that interrupts the agent mid-sentence. Two limits worth knowing: the word threshold gates barge-in only on the transcription (cascaded) stacks — on the realtime audio model the adaptive interruption detector owns barge-in instead, so the level still tunes it but the exact word count does not apply — and setting response_delay_seconds takes over the silence window entirely, so the level then shapes only how easily the caller can interrupt, not how long the agent waits.

The level the call ran under is echoed as interruption_sensitivity on the call/conversation config snapshot (null = the platform default).