Skip to content

Set guardrails

Guardrails are what the agent never does. They apply on every turn, ahead of any Playbook's instructions, and no Playbook can override one. The concept is covered in Agent Behavior; this guide is about authoring them. Guardrails are edited at Behavior · Guardrails.

The three lists

List Use it for
Topics to avoid Subjects the agent should refuse to discuss: competitor pricing, legal advice, a person's own medical situation.
Claims not to make Statements about your product or company the agent must not assert: outcome guarantees, unverified statistics, regulatory claims.
Required disclaimers Language the agent must include in certain situations: "results vary", financial-advice disclaimers, licensing statements.

Each entry is one rule, stated plainly. Write the rule, not the rationale — "Never quote competitor prices" works; a paragraph about why does not add force. General guidance belongs in Standing instructions; keep only hard limits here. The free-form Additional rules box is retired: on September 15, 2026 its text moved into Standing instructions, one instruction per line.

What belongs elsewhere

A guardrail is a hard limit, not general behavior. "Keep messages short" is Personality; "don't re-qualify the lead in the booking phase" is Playbook content. The test: would you want this rule enforced in every situation, including ones you haven't thought of? If yes, it's a guardrail.

The compliance catalog

Anychat ships guardrail sets for common compliance shapes — healthcare claims, franchise disclosure, PII handling. Importing one merges its entries into your lists, where you can edit or remove individual rules. For a regulated vertical, start from the catalog set and add your brand's specifics on top.

Built-in guardrails

Underneath your lists, the platform applies baseline rules to every Anychat agent: don't generate harmful content, don't impersonate a human, don't leak internal identifiers. These are not editable and don't appear in your lists; your guardrails add to them.

When you need enforcement, not instruction

A Guardrail is read by the agent and followed with judgment. For something Anychat must enforce rather than the agent follow, write a Rule instead: Rules are checked against the agent's actual reply, every time, whether or not the model remembered the instruction.

What's next