> ## Documentation Index
> Fetch the complete documentation index at: https://www.bolna.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# The agent interrupts or talks over the caller

> Fix a Bolna agent that cuts callers off, or won't stop talking when interrupted, by tuning endpointing, linear delay and the interruption threshold.

Two opposite complaints share one set of settings, all on the [Engine tab](/docs/agent-setup/engine-tab).

## The agent cuts the caller off mid-sentence

The agent decided the caller had finished speaking when they had not — usually because they paused to think, or read out a number in chunks.

<Steps>
  <Step title="Raise Linear Delay">
    Linear delay exists specifically to absorb mid-sentence pauses. Try **400–500 ms** as a starting point and increase for callers who dictate numbers or addresses.
  </Step>

  <Step title="Raise Endpointing">
    Endpointing is how long the agent waits after speech stops before responding. **200–300 ms** suits natural conversation; below that the agent will clip people.
  </Step>

  <Step title="Move off the Fast response rate preset">
    If the agent is on **Fast**, switch to **Balanced** or **Custom** and tune the two values above. Fast trades conversational safety for speed.
  </Step>

  <Step title="Re-test with a real caller, not a clean read">
    Have someone dictate a 12-digit number and a street address. Those are the cases that expose aggressive endpointing.
  </Step>
</Steps>

## The agent keeps talking when the caller interrupts

<Steps>
  <Step title="Lower the interruption threshold">
    The threshold is the **number of words** the caller must say before Bolna treats it as an interruption. A high value makes the agent stubborn; lower it so short interjections land.
  </Step>

  <Step title="Check whether background noise is being transcribed">
    A threshold set low enough to catch "stop" will also catch a TV in the background. Enable **noise cancellation** on the [Call tab](/docs/agent-setup/call-tab) before lowering the threshold further.
  </Step>
</Steps>

<Info>
  Stopwords such as **"stop"**, **"wait"** and **"hold on"** always pause the agent immediately, whatever the interruption threshold is set to.
</Info>

***

## If tuning does not settle it

* The agent may be responding slowly rather than interrupting — the caller starts a second sentence into the gap and collides with the reply. Check [Responses feel slow](/docs/concepts/latency) and the [per-call latency metrics](/docs/concepts/call-latencies).
* A long, list-heavy reply is more likely to be interrupted. Instruct the agent to answer in one or two sentences — see [Write better prompts](/docs/prompting/introduction).
* Enable **Generate precise transcript** on the Engine tab while debugging, so the transcript shows exactly what the transcriber heard at the moment of the collision. See [Capture precise transcripts](/docs/customizations/capturing-precise-transcripts).

***

## Next steps

<CardGroup cols={3}>
  <Card title="Tune interruptions and speed" icon="gear" href="/docs/agent-setup/engine-tab">
    Every setting on one page
  </Card>

  <Card title="Responses feel slow" icon="gauge-high" href="/docs/concepts/latency">
    If the real problem is delay
  </Card>

  <Card title="Read latency metrics" icon="chart-line" href="/docs/concepts/call-latencies">
    Per-call timings
  </Card>
</CardGroup>


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.