Skip to main content

Agents

Style and guardrails

Stopping the model improvising into claims you cannot keep.

What this field does

Style and guardrails is a free-text field in the Behavior tab of a Finn. It sits between Identity Text and Response Guidelines. Identity says who the agent is. Style and guardrails says how it speaks and what it must never say or do.

The model will fill any gap you leave. If you do not tell it never to quote a price, it will quote a price. If you do not tell it what to do when a caller gets angry, it will improvise something. Every guardrail you write is a claim the agent stops making on your behalf.

Set it in the dashboard: Create New Finn (or View All Finn → [Finn] → Edit) → Behavior tab. There is no API field for this in the docs as they stand.

What to write

Two halves. Style, then limits.

Speak in short, conversational sentences. Sound warm but professional.
Never give medical advice. Never quote prices — say "let me have someone
call you back with pricing." If the caller seems upset, slow down and
acknowledge their frustration before continuing.

Style is tone plus sentence shape. Pick three tone adjectives and stop. "Be helpful" means nothing to the model. "Warm, brief, never pushy" produces measurably different output.

Limits are the "never" rules. Each one needs a fallback, or the agent hits the wall and has nothing to say. Never quote a price on its own leaves an awkward pause. Never quote a price — transfer to the front desk if asked gives the agent somewhere to go.

The guardrails you almost certainly need

GuardrailWhy it matters
Disclose AI status when asked directlyDeceiving a caller who directly asks if they are talking to a human is illegal in most jurisdictions. See compliance.
Honor opt-out immediately and end the callOpt-outs must reach your suppression list. See compliance.
Never quote a price or a figure not in the knowledge basePriced claims are the ones customers hold you to.
Never state availability the agent cannot verifyThe agent will invent a Thursday slot if you let it.
Never give medical, legal, or financial adviceRegulated speech. Route to a human.
Escalation path after N failed clarificationsWithout it, the agent loops. Two attempts then offer transfer works.

Recording disclosure is a separate matter and belongs in the welcome message, not here. See agents-welcome and recording-consent.

Positive and negative together

A field that is only "never" rules gives the agent no frame to work inside. It becomes evasive because avoidance is the only behavior you described. Pair each prohibition with the thing it should do instead.

Rules:
- Always ask "Am I speaking with {customer_name}?" before sharing any
  appointment details.
- If the caller says "remove me" or "stop calling", honor immediately and
  end the call politely.
- If asked if you're a real person, say honestly: "I'm Aria, a voice
  assistant for Smile Dental."
- Never quote a price — transfer to the front desk if asked.
- Never schedule a Sunday appointment — the clinic is closed.
- If the caller's intent is unclear after two clarifying questions, offer
  human transfer.

Guardrails read better as conditionals than as adjectives. If X, do Y survives pressure. Be careful about X does not.

What goes wrong

SymptomLikely causeFix
Agent quotes a number you never approvedNo rule covering that topicAdd the rule plus a fallback line
Agent refuses everything, sounds cageyNegative-only guardrailsAdd the positive counterpart to each never
Agent still breaks a rule you wroteRule stated as tone, not as a conditionRewrite as If X, say Y
Tone drifts late in long callsGuardrails carry rules but no examplesAdd 2–4 micro-dialogues under Behavior. See agents-prompting
Rules contradict each otherSame constraint stated in Identity Text, Style and guardrails, and Response GuidelinesKeep each constraint in exactly one field

That last one is the common failure. Identity Text, Style and guardrails, and Response Guidelines all feed the same model. Duplicate a rule across two of them with slightly different wording and you have given the model a choice it will resolve unpredictably.

Facts do not belong here

Guardrails describe behavior. Business facts (hours, pricing tables, policy text, product specs) belong in the knowledge base, where you can update them without touching the agent's behavior. See agents-knowledge-base.

The guardrail field should reference the boundary, not restate the content. Never state a price that is not in the knowledge base is a guardrail. A price list pasted into the guardrail field is a maintenance problem.

Verifying it holds

Guardrails are unverified until a hostile caller tests them. Use Test on Chat for fast iteration and Test on Call for the real thing. See agents-testing.

Test the cases that break agents:

  • Ask directly whether it is an AI.
  • Ask for a price.
  • Say "take me off your list."
  • Give a deliberately ambiguous answer twice in a row.
  • Get angry.

When a call goes off-script, listen to the recording rather than guessing. Find the first turn where behavior diverged. It is usually a missing rule, not a model failure. Add a rule for that exact scenario, then re-test. Recordings and transcripts are in call-logs.

Changes apply to new calls only. In-flight calls finish under the old configuration. Configuration history is kept, so a guardrail change that degrades behavior can be rolled back. See agents-versioning.