Appearance
Guardrails
Guardrails are the explicit instructions and limits that apply across an entire call: what the agent must never say or promise, which topics it declines and hands off instead, and when it must stop and transfer to a person.
Keeping guardrails separate from the conversation flow is deliberate. Rules you can review on their own, without reading the whole workflow, are rules you can actually trust are being followed.
(No product screenshot is available for the guardrails panel yet — the fields and layout below are described at a concept level rather than illustrated.)
What to put in guardrails
- What the agent must never promise — discounts, delivery dates, guarantees, legal or medical advice, anything you can't actually back
- Topics it should decline — and what it should do instead (usually: offer a human handoff)
- Hard stops — situations where the agent must end the call or transfer immediately
- What it says when it doesn't know something — an honest "I don't have that, let me connect you" beats a confident guess
Write them as rules, not conversation
A guardrail is a constraint the agent checks against, not something it says out loud verbatim. Write it as an instruction ("never quote a price without confirming with a human") rather than as example dialogue.
Test that they actually fire
A guardrail you haven't tested is a guardrail you're assuming works. In Browser Test, deliberately try to get the agent to break each guardrail — ask for a discount it shouldn't give, raise a topic it should decline, ask for a human. If nothing happens, the guardrail isn't doing its job yet.