Voice agents
Model and prompt
Choose the language model behind an agent and configure its prompt, greeting and generation settings.
Choosing a model
The model decides what the agent says and when it uses tools. On the phone, speed matters as much as intelligence: the caller hears silence until the first token arrives.
| If you need | Start with |
|---|---|
| Fast, reliable front-desk conversations | A fast general model such as GPT-4o or a Claude Sonnet-class model |
| Complex reasoning over long policies | A larger model, and keep replies short in the prompt |
| Lowest cost at volume | A small model, with a tightly scoped prompt and good knowledge |
The model catalog in the dashboard (Models) shows the models enabled for your workspace. If you bring your own OpenAI or Anthropic key, calls on those models use your key. See Bring your own keys.
Settings
- System prompttextrequired
- Who the agent is and how it behaves. At least 20 characters.
- First messagetext
- What the agent says when the call connects.
- First message modeenum
assistant-speaks-first(default),assistant-waits-for-user, orassistant-speaks-first-with-model-generated-message.- Temperaturenumber
- 0 to 2. Default 0.7. Lower is more consistent; 0.3 to 0.6 suits most phone agents.
- Max tokensinteger
- 50 to 4000. Default 250. Caps the length of each reply.
- Knowledge basereference
- Documents the agent can search during the call.
- Emotion recognitionboolean
- Lets the model take the caller's tone into account.
Tip
Keep max tokens low for voice. A 250-token cap stops long monologues, which callers hate and which cost more speech time.
Writing the system prompt
A good voice prompt has four parts: identity, goals in order, facts, and style rules. See the Prompting guide for patterns and anti-patterns.
Last updated November 2026 · Found a problem? Email support@katexs.com
