There are several ways and parameters to configure how your agent is going to interact with the user — when to backchannel, when to interrupt, how responsive it is, etc.

Who Speaks First

You can configure whether the agent speaks first or the user speaks first, no matter if you are using Retell LLM framework or your own LLM.

Retell LLM Framework

You will play with the begin_message field in Retell LLM API to control who speaks first.

  • Agent speaks first with LLM generated (based on your prompt) message: leave begin_message unset.
  • Agent speaks first with a specific message: set begin_message to the message you want agent to speak.
  • User speaks first: set begin_message to empty string.

Custom LLM

If you are using custom LLM, then you need to set the content of the first response event (the one with response_id = 0) you send.

  • Agent speaks first: set content to the message you want agent to speak.
  • User speaks first: set content to empty string.

Backchannel

Backchannel is the ability for the agent to make small noises like “uh-huh”, “I see”, etc. during user speech, to improve engagement of the call. You can set whether to enable it, how often it triggers, what words are used.

You will look at the following fields in Create Agent API:

  • enable_backchannel: set to true to enable backchannel. Default to false.

  • backchannel_frequency: When our engine determines that a backchannel is possible and appropriate here, you don’t necessarily always want to backchannel. Humans often only backchannel at some places, not all possible cases. Therefore, this field controls how often the backchannel is triggered, with 0 being never, 1 being always when possible. The default is 0.8.

  • backchannel_words: The words that the agent can use as backchannel. The default is different for each language, and for each voice provider, as some voice provider does not have support for some of the words. The following tab contains the default backchannel words for each language and voice provider. You can overwrite this field to customize the backchannel words. Note that certain voices do not work too well with certain words, so it’s recommended to expeirment before adding any words.

    • 11labs voices: [“okay”, “uh-huh”, “mhmm”, “yah”]
    • openai voices: [“okay”, “uh-huh”, “yah”]
    • deepgram voices: [“okay”, “uh-huh”, “yah”]

Reminder

Reminder is the ability for the agent to remind the user when the user is unresponsive. You can set whether to enable it, how often it triggers, and how many times it can trigger.

You will look at the following fields in Create Agent API:

  • reminder_trigger_ms: The time in milliseconds after the user is unresponsive that the reminder is triggered. The default is 10000 (10s).
  • reminder_max_count: The maximum number of times the reminder can trigger. The default is 1. Set to 0 to disable agent from reminding.