Agent settings
Configure your agent: system prompt and guardrails, model and temperature, status (draft, active, archived), languages, fallback behavior for "I don't know", escalation, and the public appearance of the chat widget.
View as MarkdownAn agent's settings control how it answers — its instructions, the model behind it, what it does when it doesn't know, and how it looks to customers. Knowledge comes from data sources; everything else lives here, in the agent's Settings.
System prompt
The system prompt is your agent's standing instructions — its persona, tone, scope, and rules. It's prepended to every conversation. A good support prompt does three things: sets the voice, defines the scope, and draws hard guardrails.
You are the support assistant for Acme, a store selling outdoor gear. Be friendly, concise, and never pushy. Answer ONLY from the provided knowledge. If something isn't covered, say you'll connect the customer with the team — never invent policy, prices, order details, or shipping dates.
When you create an agent you can describe it in plain language and the agent builder drafts a starting system prompt for you. Edit from there rather than starting from a blank box.
Explicitly tell the agent what to do when it doesn't know and what it must never do. Without a guardrail, a model can drift — paraphrasing "30-day returns" into "no-questions-asked refunds." See Best practices.
Model and temperature
Each agent runs on one model, chosen from your model picker. The model determines answer quality, speed, and the credit cost per reply.
Temperature controls how deterministic the model is, on a 0 to 1 scale. For support, lower is better.
| Temperature | Behavior | Use for |
|---|---|---|
| 0.0 – 0.3 | Focused and consistent; sticks closely to the retrieved facts. | Most support agents. Recommended. |
| 0.4 – 0.6 | Balanced; a little more natural phrasing variation. | The default range; fine for conversational tone. |
| 0.7 – 1.0 | Creative and varied; higher chance of drifting from the source. | Rarely a good idea for grounded support. |
For a deeper comparison of models and how to choose, see Models & model choice.
Agent status
An agent moves through three states:
| Status | Meaning |
|---|---|
| Draft | Being built and tested. Visible in the playground; not yet treated as live. |
| Active | Live and answering on its connected channels. |
| Archived | Retired. Kept for its history but no longer serving customers. |
Set an agent to Active when you're ready to deploy. You can keep refining its data and prompt while it's active — changes go live immediately.
Fallback behavior
The fallback controls what happens when retrieval finds nothing relevant. A well-configured agent admits it doesn't know and offers a path forward rather than guessing. You can set:
- Fallback message — what the agent says when it can't answer from your knowledge.
- Lead form — optionally collect the customer's details when the agent can't help, so your team can follow up.
- Handoff — route the conversation to a human via the Escalate to a human action, landing it in your Help Desk.
Pair a clear fallback message with escalation. "I'm not sure about that — let me connect you with our team" plus a handoff is far better than a confident guess.
Languages
Set the languages your agent should support. The agent can detect and reply in a customer's language, and language settings also drive Help Desk translation for human-handled tickets.
Appearance and public settings
Public settings control how the agent looks and greets customers on the website widget:
- Brand color — the accent color of the chat bubble and header.
- Avatar — the agent's profile image.
- Welcome message — the greeting shown when the chat opens.
- Suggested prompts — starter questions shown as one-tap chips to nudge customers toward what the agent does well.
These apply to the website widget and other surfaces. For the embed itself, see Website widget.