Progress and error messages

Note: Progress messages, error messages, and empty LLM response handling all apply to regular (non-speech-to-speech) models. For speech-to-speech models, see Inactivity reminder. Max turns applies to both.

Progress message

Sometimes it may take the Agent too much time to generate response, for example, due to use of multiple tools or slow tool response. In such cases, consider using 'Progress message' and 'Progress timeout' parameters in the Miscellaneous tab to define progress message that will be played to user while agent is “thinking”.

In addition to the time-based progress message described above, you can configure progress messages per tool — played before, on success, or on failure of that specific tool's execution — on the tool's Progress tab. See Tool progress messages.

Alternatively, you can play progress messages when specific tools are called using the legacy progress_message_conditions advanced configuration parameter. For details, see Conditional progress messages when calling tools.

How progress messages interact with the model's own speech

An agent has three ways to keep the user engaged while it works, and the framework coordinates them so it never talks over itself:

  1. the time-based progress message above (played when the whole turn is slow);
  2. per-tool messages — During execution / On success / On failure (see Tool progress messages);
  3. dynamic per-tool progress messages — the agent generates that tool's During execution line itself, based on the state of the conversation (see Letting the model write the message).

When the model speaks its own message alongside a (non-terminal) tool call, that message already fills the pre-tool wait, so the framework suppresses the redundant pre-tool reassurance for that step:

Everything else is unaffected: the tool's On success / On failure messages still play (they report the outcome, not the wait), and a different tool called later in the same turn — one with no accompanying message — still plays its own During execution message. Per-tool messages can play more than once per turn; the time-based message plays at most once.

Whether the model's message actually reaches the user on a non-streaming connection is controlled by the llm_message_with_tool_call advanced configuration parameter (play delivers it, drop discards it). With TTS streaming it is always delivered. If the message is dropped, nothing filled the wait, so the pre-tool reassurance is not suppressed.

Call-control tools. end_call and transfer_call are the exception. Any message the model wants spoken before hanging up or transferring comes from the tool's own termination_message / transfer_message parameters (not from llm_message_with_tool_call), and these tools end or transfer the turn — so they don't use the Before / On success / On failure progress scheme.

Pause after a progress message

On voice calls, when the agent's response is ready while a progress message is still playing, the response waits for the progress message to finish and then for a short pause, so that the two don't run together. The pause is 500 msec by default; change it with the progress_message_pause advanced configuration parameter:

Parameter Type Description
progress_message_pause int Pause in milliseconds between a progress message and whatever the agent plays after it.
Optional, default is 500, range is 0 to 5000. 0 disables the pause.

For example:

{
    "progress_message_pause": 800
}

The pause follows any kind of progress message – time-based, per-tool or conditional – and applies to whatever is played next, another progress message included.

Error message

If Live Hub AI Agents Framework fails to receive valid response from LLM or encounters some other unexpected error, it generates “error response” to user. The default value for this response is “Sorry, I didn't get it. Can you please say it again?”

You may change the value by configuring ‘Error message’ parameter in the Miscellaneous tab, in the Agent configuration screen. For example:

Sorry, can you please rephrase your last question?

Max turns

When a conversation exceeds the agent's ‘Max utterances’, the agent does not hang up straight away. It sends the model a reminder, asking it to wrap up the conversation and end the call itself – for example, by saying goodbye and hanging up. If the model is still going grace_turns responses later, the call is ended for it.

Customize this behavior with the max_turns_reminder advanced configuration parameter:

Parameter Type Description
text str Reminder text sent to the model.
Optional, defaults to NOTE: This conversation has gone on for too long. Politely wrap it up and end the call.
grace_turns int Number of model responses, counting the response to the reminder itself, within which the model should end the call.
Optional, default is 2, maximum is 10. 0 ends the call as soon as the limit is exceeded, without a reminder.

For example:

{
    "max_turns_reminder": {
        "text": "NOTE: The call is running long. Offer to help with anything else briefly, then say goodbye and end the call.",
        "grace_turns": 3
    }
}

To customize the response provided to the user when the call is ended for the model, specify the max_turns_message advanced configuration parameter. It replaces the model's last response, and is not supported for speech-to-speech models:

{
    "max_turns_message": "Sorry, I need to hang up."
}

When the agent hands the conversation over to another agent during the grace period, the grace period continues – the next agent does not get a fresh ‘Max utterances’ allowance. A response given by a consulted sub-agent counts as a turn of the agent that consulted it. Amazon Nova Sonic v1 cannot receive the reminder, so it ends the call as soon as the limit is exceeded.

Note: Flows do not support max_turns_reminder. A flow is closed as soon as it exceeds its ‘Max utterances’, with max_turns_message as its response.

Empty LLM responses

Some LLMs may produce empty responses due to various technical issues or intermittent processing errors. By default AI Agents perform automatic retry upon empty LLM response, thus ensuring that conversation continues smoothly and user experience is not disrupted. You may change this behavior by configuring empty_llm_response advanced configuration parameter.

Parameter Type Description
empty_llm_response enum Configures how the AI Agent treats empty LLM response.
Supported values:
• retry (default) – retry response generation; if response is still empty after the retry, treat it as an error
• ignore – treat empty response as an error
• allow – treat empty response as a valid response