Guardrails

Writing the rules an agent must keep: the Guardrails field, what belongs in it, and how to check that it holds.

Single-prompt agents

Guardrails is the third prompt field on Settings: style guidelines, phrases to use or avoid, and the boundaries of what the agent will do. It is where you put the rules a compliance or support lead would insist on, and it is read on every turn alongside Agent Identity and Tasks.

The Guardrails textarea with a list of rules — The Guardrails field.

What belongs in Guardrails

Kind of rule	Example
Topics to refuse	Never give medical, legal or financial advice. Offer to book a consultation.
Facts not to invent	Do not quote prices, dates or availability you were not given. Say you will check.
Identity	Do not claim to be a person. If asked, say you are a virtual assistant.
Data handling	Do not ask for card numbers or passwords. Direct the caller to the secure link.
Escalation	When the caller asks for a person, or is upset, use the transfer tool.
Style	One question at a time. Replies under two sentences. Read digits one by one.
Language	Answer in the caller’s language if it is one you support; otherwise say so.

Writing rules that hold

One rule per line, most important first.
Say the alternative. “Never X” on its own leaves the model without a next move; “Never X. Instead, Y.” gives it one.
Be concrete. “Do not discuss competitors” is vaguer than “If asked about other clinics, say you can only speak for Riverside Dental”.
Do not repeat the tasks. Guardrails are limits, not the plan. Duplicating the task list here dilutes both.
Keep it short. A long list of rules is followed less reliably than a short one; move style notes into Agent Identity if the list grows.

Controls outside the prompt

Tools: an agent can only take actions it has tools for. No transfer tool, no transfer. See Tools overview.
Outcomes: capture whether a rule was broken as a field, for example gave_medical_advice with values yes and no, and filter Call Logs on it. See Outcomes.
PII redaction and retention protect data after the call regardless of what was said. See Privacy and data.
Call duration and silence limits end calls the prompt cannot. See Silence, reminders and call duration.
Evaluations replay scenarios against every version and flag regressions. See Evaluations.

Verify

Write a test scenario for each rule under Evaluation: a caller who asks for a diagnosis, a caller who asks for a person, a caller who asks for a price. Run them after every prompt change.