Agent responded to itself

Support Help
1

My agent asked the question it was meant to ask, then immediately started responding to its own question in a new message.

Could I get the full address for the project, including the ZIP code? It also helps if you spell out the street name for me.

It’s one two three four five, S-

and then got cut off by the caller reading out their actual address.

Call ID: call_00a7cbab2389ae9f706692b6651

2

Hey @zachsents I have escalated your Call ID to the team for a review.

Thank You

3

Hello @zachsents Looked into call_00a7cbab2389ae9f706692b6651. Both messages were emitted by the agent. The second was triggered by the caller’s overlapping “Yep.” being treated as a completed end-of-turn event, which dispatched a fresh LLM call with effectively no usable user input. The model then produced a templated synthetic response in the shape of the digit-spelled address exemplars in your prompt, and TTS spoke it before the caller could provide the real address.

A few changes that should prevent this:

  1. Replace the fully-formed example addresses in your Collect Address node and global Speech section with schematic placeholders (e.g. <digit digit digit digit> <Street Name spelled out>). When user input is sparse, smaller cascading models tend to anchor to and reproduce concrete exemplars verbatim.
  2. Lower interruptionSensitivity from 0.8 toward 0.6 so single-word backchannels like “Yep.” during the tail of an agent turn are not treated as completed user turns.
  3. Consider pinning the Collect Address node to a stronger model. You are already using a stronger model on your Welcome and Qualify nodes, and the cascading default is the most likely producer of this behavior.

The call completed normally and the address was ultimately collected correctly, so this was a single-turn glitch rather than a call failure.

Thank You

4

I don’t believe we have that exact example in our prompt that the agent tried to copy. Unfortunately, we have interruption sensitivity high because we’ve been dealing with some latency concerns. We’ll leave it for now and chalk it up to AI weirdness. Thanks for checking into it!