Agents
Agent settings reference
Every field on an agent's Settings tab: model, key, temperature, max tokens, output format, sampling, token prices, rate limit and status.
Reference2 min readUpdated 23 October 2026
The Settings tab holds everything about how an agent generates its answers. Changes reach every channel (API, widget, workflows) within seconds; agent configuration is cached for up to five minutes and the cache is cleared when you save.
Fields
| Field | Default | Range and notes |
|---|---|---|
| Name | Required, up to 120 characters, unique in your account. | |
| Description | Optional. | |
| Provider | OpenAI | OpenAI, Anthropic, Google Gemini or Groq. |
| Model | gpt-5.5 | Chosen from the provider's current list. A retired model already on an agent stays visible. |
| API key | Your provider key, stored encrypted with AES-256-GCM and shown masked. Pick a saved key or paste a new one. | |
| Temperature | 0.7 | 0 to 2, in steps of 0.1. Lower is more consistent; higher is more varied. |
| Max tokens | 4,096 | 1 to 128,000. The most tokens a reply may use. Your model may allow fewer. |
| Output format | text | text, markdown, JSON or code. JSON turns on the provider's JSON mode. |
| Top P | 1.0 | 0 to 1 (Advanced). |
| Frequency penalty | 0 | -2 to 2 (Advanced). |
| Presence penalty | 0 | -2 to 2 (Advanced). |
| Input / output cost per 1M tokens | 0 | In US dollars, used only for the Cost tab's estimates (Advanced). It doesn't change what you're billed. |
| Rate limit | 60 requests a minute | Read only. Shared by every caller of the agent: API, widget and workflow Agent steps. |
| Status | Active | Active or Paused. A paused agent turns every request away. |
Behaviour you can't switch off
- Knowledge is used automatically once the agent has indexed content. There is no knowledge switch.
- Grounding: agents with knowledge get a fixed instruction added to their prompt that tells the model not to make things up and to answer from the passages it was given.
- Tool rounds: an agent with MCP tools can call them for at most 8 rounds per reply, each call limited to 10 seconds. After that it must answer.
- Time limit: a reply must finish within 120 seconds or it ends with a
TIMEOUTerror event. - History: the last 12 messages of the conversation, up to 12,000 characters, go to the model with each new message.
Retries
If a provider is overloaded, SentientOne retries before giving up: Anthropic and OpenAI-compatible providers three times on 529 or 503, Google three times on 429 or 5xx, and Groq's tool_use_failed twice with a lower temperature. There is no automatic switch to a different model.
Per-request overrides
From the API you can replace the system prompt for one request with the prompt field, use a different provider key with the X-Llm-Key header, and let a paused agent answer with allow_unpublished: true. See Chat stream.
Still stuck?
Email us with what you tried and any request id. A person replies.