Think

The Think section controls the conversation: the system prompt, the greeting, the language model (LLM), and what the agent does when a caller goes quiet. For how Listen, Think, and Speak fit together, see the Execution Layer overview. For how to write the prompt itself, see Agent behavior.

Prerequisites

Before you begin:

An agent already exists. If not, see Create an agent.
These steps change a live agent. The API path uses PATCH; see Update an agent.

Set the prompt and greeting

Dashboard
API

Open your project and go to the Think section. Edit the System prompt and the Greeting the agent opens with.

The Think section with the system prompt and greeting fields.

system_prompt and greeting are top-level fields, so you can update them on their own. Set inbound_greeting or outbound_greeting to open phone calls with a different line.

{
  "system_prompt": "You are a support agent for Wayne Enterprises.",
  "greeting": "Hi, how can I help you today?"
}

Change the language model

Dashboard
API

Open your project and go to the Think section. Open the LLM model dropdown and pick a model. As with Listen, the dropdown separates SLNG models, hosted in your region, from BYOK, models you reach with your own provider key.

The Think section with the LLM model selector — The Think section LLM model selector.

To use the BYOK tab, register your provider key first; see Bring your own key.

Send the whole models object. A PATCH replaces it as a unit rather than merging fields, so change llm and keep the other model fields as they are.

{
  "models": {
    "stt": "deepgram/nova:3",
    "llm": "bedrock-mantle/nvidia.nemotron-super-3-120b:latest",
    "tts": "slng/fish/tts:s2.1-pro",
    "tts_voice": "16cabdb7f8d240569aff36c9e480d783"
  }
}

Prefer a model under the slng/ prefix to run in your region. A bare provider/model uses the provider’s model through SLNG; add your own key (BYOK) to run it on your account. See Where a model can run.

To see which models your region and language support, see Which models are available.

Tune generation

models.llm_kwargs sets the generation parameters for each turn. The ones worth setting for a voice agent:

temperature: lower keeps answers consistent, higher makes them varied.
max_completion_tokens: caps how long a single reply can run.
top_p: nucleus sampling, an alternative lever to temperature.
tool_choice and parallel_tool_calls: how the model reaches for tools.

{
  "models": {
    "stt": "deepgram/nova:3",
    "llm": "bedrock-mantle/nvidia.nemotron-super-3-120b:latest",
    "tts": "slng/fish/tts:s2.1-pro",
    "tts_voice": "16cabdb7f8d240569aff36c9e480d783",
    "llm_kwargs": { "temperature": 0.7, "max_completion_tokens": 4096 }
  }
}

Handle silences

idle_nudges sets what the agent says when a caller stops responding. It nudges once, nudges again, then ends the call, each after its own delay. Edit the text and delays, or set enabled to false to keep quiet.

{
  "idle_nudges": {
    "enabled": true,
    "first_nudge_delay_seconds": 15,
    "first_nudge_text": "Are you still there?",
    "hangup_delay_seconds": 15
  }
}

Route with the Context Router

The Context Router sits in front of the language model and decides how to handle each turn, so some turns skip the large model entirely. Toggle it with llm_router_enabled. For how routing works, see Context Router.

Connect tools and variables

Tools and MCP servers are attached separately, not set in this section, and tool_refs and mcp_refs cannot be changed with PATCH. See Choose a tool.

Keep Think reliable

The LLM step can fall back to another model and has its own per-turn timeout (models.fallbacks.llm, models.llm_first_token_timeout_s). These work the same way across all three sections. See Reliability.

Next steps

Speak: the text to speech model and voice.

Get started

Regions

Execution layer

Agent Builder

Models

Integrate

Prerequisites

Set the prompt and greeting

Change the language model

Tune generation

Handle silences

Route with the Context Router

Connect tools and variables

Keep Think reliable

Next steps

​Prerequisites

​Set the prompt and greeting

​Change the language model

​Tune generation

​Handle silences

​Route with the Context Router

​Connect tools and variables

​Keep Think reliable

​Next steps

Prerequisites

Set the prompt and greeting

Change the language model

Tune generation

Handle silences

Route with the Context Router

Connect tools and variables

Keep Think reliable

Next steps