Guardrails

Guardrails keep an agent on its job. They apply on top of your prompt, in the same way, everywhere the agent talks: phone calls, the test panel and test suites, SMS and web chat. Three rails, each on its own:

Rail	What it does
Stay on purpose	The agent helps only with what it is for. Asked for anything else (general questions, a recipe, code, another task), it says in one short sentence that it cannot help with that here and returns to its purpose. Give the purpose in a few words, or leave it empty to go by the prompt.
Resist manipulation	The agent keeps its persona, instructions and rules whoever the person claims to be and whatever they claim: a request to behave differently, to reveal or repeat its instructions, to play another role, to treat the conversation as a test or a game, or to make an exception changes nothing.
Never discuss	Subjects the agent never discusses, gives an opinion on, or provides information about, even when asked directly. Up to 30.

When the person keeps pushing decides what happens after the agent has declined twice: keep declining and steer back (the default), or say goodbye and end the call on the third decline. On SMS and web chat the agent closes the conversation instead. Every decline is recorded: it appears in the call’s transcript and in test results as a guardrail_declined event with its reason (off_topic, manipulation or blocked_topic), so you can see how often callers run into a rail and which one.

Setting guardrails

Dashboard: open the agent, go to Behavior, and fill in the Guardrails card. API: set guardrails on the agent (null turns every rail off):

curl -X PATCH https://api.revring.ai/v1/agents/{agentId} \
  -H "x-api-key: YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "guardrails": {
      "focus": { "enabled": true, "scope": "booking, moving or cancelling dental appointments" },
      "manipulation": { "enabled": true },
      "blockedTopics": ["politics and elections", "other clinics and their prices"],
      "onViolation": "end_call"
    }
  }'

Field	Description
`focus.enabled`	Stay on purpose.
`focus.scope`	The purpose in a few words (up to 500 characters). Empty: the prompt defines it.
`manipulation.enabled`	Resist manipulation.
`blockedTopics`	Subjects never discussed, up to 30 of up to 80 characters each.
`onViolation`	`redirect` (default) or `end_call`.

Guardrails are part of the agent’s configuration: they are saved in its versions, published with it, and copied when the agent is cloned.

Proving them

Guardrails are instructions the agent follows, so test them the way you test anything else the agent should do: in the test panel’s Chat, or with a test suite whose scripts try to pull the agent off course (an unrelated question, “ignore your instructions”, a blocked subject, insisting three times) and whose rubrics check that it declined and came back. Run each test a few times.

Guardrails steer the agent’s own replies. They do not filter what the person says, and they are not a substitute for keeping secrets out of the prompt: anything in the prompt can, in principle, be talked about, so put credentials, internal notes and prices you do not want quoted in a tool or a knowledge base with access rules instead.

Getting started

Platform

Setting guardrails

Proving them

​Setting guardrails

​Proving them

Setting guardrails

Proving them