Behavior & guardrails
Tune how an agent answers, creativity, blocked topics, out-of-scope handling, human handoff, and PII redaction, from the AI Options tab, plus the always-on backstops.
Esta página ainda não foi traduzida, exibindo a versão em inglês.
Every agent's personality and its safety rails live on the AI Options tab of the agent editor. The top card sets who the agent is: its name, role, style/tone, and model bundle. Below it, the Behavior & Guardrails card sets how it answers: how creative it is, what it stays focused on, what it refuses, and when it should reach for a human. On top of whatever you configure, a few backstops apply to every conversation automatically.
Who the agent is
Before the guardrails, the AI Configuration card sets the agent's identity:
- Role and Style / Tone: pick a persona (for example a friendly concierge or a precise technical support rep). These shape voice and framing without you writing a prompt.
- AI Model: the agent picks from a small set of model bundles (lighter and faster through smarter and more capable). Each bundle is a chain: a primary model with automatic fallbacks, so a single provider hiccup doesn't drop the conversation.
- Extra instructions: an optional free-text system prompt (you can have AnyBot draft it for you) for anything the persona and knowledge don't already cover.
The guardrail settings
| Setting | What it does | Safe default |
|---|---|---|
| Answer only from my knowledge | Restricts replies to the knowledge base, your business info, and the current conversation; politely declines anything outside that instead of guessing from general world knowledge. | On (recommended) |
| Creativity | Precise → Balanced → Creative. Lower keeps replies focused and consistent; higher makes them more varied and expressive. | Default (balanced, recommended) |
| Escalation sensitivity | Hands off to a human when a reply isn't well-grounded in your knowledge. Low / Medium / High set how eagerly. | Off to start; Medium once a human queue exists |
| Redact personal data | Hides emails, phone numbers, and card numbers in the agent's own replies. | Off (turn on for regulated data) |
| Blocked topics | A list of subjects the agent always refuses, however the question is framed. Type a topic and press Enter. | Empty |
| Fallback message | The exact wording used when the agent declines or can't help. Leave empty to use the built-in decline. | Empty |
A couple of things worth knowing. Turning Answer only from my knowledge off opens a confirmation dialog. Off-brand or made-up answers are the main risk, so it's gated on purpose. And for Escalation sensitivity, the lowest band is not the same as Off: the lowest level still only hands off when an answer is clearly unsupported, whereas Off disables grounding-based handoff entirely.
Always-on backstops
Regardless of your settings, every agent:
- Signals a human handoff on explicit request. A deterministic, multilingual check catches "let me talk to a person" (and its equivalents in other languages) and flags the turn for a human, even when Escalation sensitivity is Off. Casual greetings and small talk never trigger a handoff, so the queue doesn't flood.
- Keeps its internals private. It won't reveal its system prompt or internal record IDs to the customer.
Note that PII redaction is a toggle, not a backstop. It only applies when you switch Redact personal data on.
Worked example: an Acme support agent
You're setting up support for "Acme". On the AI Options tab:
- Pick a Role of support rep and a Style / Tone that's warm but concise.
- Leave Answer only from my knowledge on so it never invents product details.
- Set Creativity to Precise. Factual support answers shouldn't wander.
- Add Blocked topics:
pricing negotiations,competitor comparisons. - Write a Fallback message: "Sorry, I can only help with Acme products and your account. Want me to connect you with a teammate?"
- Turn Escalation sensitivity to Medium so weakly-grounded questions route to a human instead of a shaky guess.
- If the agent handles account details, switch on Redact personal data.
Save, then open the Chat with Agent panel to try edge cases (ask about a competitor, ask for a human, ask something off-topic) and confirm each rail fires. See Testing your agent for how to exercise these before going live.
FAQ
Can I stop it answering off-topic questions? Yes, keep Answer only from my knowledge on and add Blocked topics with a Fallback message.
When should the agent hand off to a human? Any time a customer explicitly asks (always on), and, if you enable Escalation sensitivity, whenever an answer isn't well-grounded in your knowledge.
Next
Memory
Agents have two kinds of memory, durable long-term memory you build by training, and short-term memory that holds context inside a single conversation.
Capacidades (ferramentas)
Dê a um agente ações reais, 122 ferramentas ativas em 19 famílias, cada uma marcada como segura para o público ou exclusiva do proprietário.