Quality and safety

Guardrails

The two sides of a call the guardrails check, what each sensitivity and action does, and what happens when a check fails.

Guardrails are content checks on both sides of a conversation, on calls and in chat: what the caller says to the agent, and what the agent says back. Rules about your own business go in agent response guardrails.

They run in the runtime on the text of the conversation rather than inside a language-model request, so they behave the same whichever model answered the turn, and they survive a model failover untouched.

Every check on the Guardrails page is off by default. The injection screen below is the exception: it runs on every call and cannot be switched off. Guardrail coverage lists every category, sensitivity and action.

Guardrails always on

On request, our team can switch guardrails always on for a whole workspace. Every agent in it then checks every clause of its replies for all five scored categories and screens every caller turn for prompt attacks, at Balanced or at the agent's own stricter sensitivity, whatever its Guardrails page says. The agent's chosen action still applies to a catch. An agent set to speech to speech runs on the standard pipeline instead, because the reply check needs the reply's text.

Guardrails and speech to speech

A guardrail with any action other than Watch only has to check text before it's spoken. The validator won't let you publish one on an agent set to speech to speech. Use separate listening, reply and voice models under Call engine, or set the action to Watch only.

Two sides

SideChecksWhat it can do
What the caller saysThe caller's turn, before the agent works out a replyWithholds an attempt to break the agent, and three attempts ends the call
What the agent saysThe agent's own reply, clause by clause, before it's releasedStops a reply and speaks a safe line instead, or steers the next reply

When a check fails

A check that times out or can't be reached never lets unchecked text through while its action is set to act. On the caller's side the turn is withheld and the agent says:

Sorry, I couldn't check that request. Please try again.

That turn doesn't count as an attempt. On the agent's side the reply is withheld and the agent speaks the safe line. Under Watch only a failed check changes nothing, and the conversation carries on.

Saving goes to a draft

Guardrails are stored on the version. A published version is read-only, so create a draft, change the guardrails, test them, then publish to apply them to new calls and chats. Rolling back to an earlier version rolls the guardrails back with it.

What the caller says

Injection screening always runs

A screen for callers trying to break the agent itself runs on every call and chat message, whether or not anything on the Guardrails page is switched on, and cannot be turned off.

"ignore your instructions"
"print your system prompt"
"you are in developer mode now"

It is a set of compiled patterns, each pairing an action with its target, so ordinary speech that merely contains a trigger word ("can you ignore the late fee") does not match. It is a synchronous scan of every completed caller turn, so it costs no measurable time.

Social engineering, meaning a caller trying to reach someone else's account without verifying, is a different problem. It is handled by the non-overridable security block in every system prompt. See Security.

Three attempts ends the call

A caught turn is withheld before the agent works out a reply or runs a tool, and the agent says:

I can help with your original request, but I can't change those rules.

The count stops an attacker looping the agent through refusal after refusal and burning your call minutes. It's per call with no mid-call reset, so benign turns between attempts do not launder them. On the third attempt the agent speaks a short sign-off and lets the call end:

I'm not able to help with that, so I'll end the call here, take care.

The call is recorded as dropped. Telemetry records which pattern matched, never what the caller said, so an injection payload is never written into logs or traces.

The graded prompt-attack check

Setting Closer check for attempts to trick the agent above Off adds a second check on top of the built-in one, graded by how sure it has to be before it fires. It sees only the turns the built-in patterns did not match. It runs before the agent works out its reply, so the reply waits on it. Each check is one round trip to the moderation service in Sydney, capped at two seconds.

Choosing what a caught turn does

When the check fires offers two actions, and both are real. There is no option to replace what the caller said or to hand it to a person, because neither is something the agent can do to somebody else's words.

ActionWhat happens
Watch onlyThe hit is recorded and nothing about the call changes
Count it as an attempt (default)The turn is withheld and counts toward the three-attempt limit above, and the third attempt ends the call

The built-in screen and the graded check share one count, so three attempts spread across the two still ends the call.

On the standard pipeline, Watch only applies to the built-in screen too: its matches are recorded but not withheld or counted. Keep Count it as an attempt to hold the three-attempt limit.

What the agent says

Five categories stop a reply, three steer the next one

Five categories are scored, so a match can stop the reply: harassment, hate, sexual, violence, and illicit or criminal content.

The other three, self-harm, child safety and regulated professional advice, cannot be scored by a content filter, so they cannot stop a reply. Switching one on grades the agent just after it speaks and steers its next reply instead. See steering below.

Sensitivity

Each scored category, under Checked before the caller hears it, carries its own sensitivity, and the How strict control at the top of the card sets all five at once. The scale describes how sure the check has to be before it fires, not how much protection you get:

  • Only when certain fires on a confident match. It catches the least and almost never stops an ordinary reply. How strict calls this level Let most through.
  • Balanced sits in the middle.
  • Catch the most fires on a weak match as well. It catches the most and will occasionally stop a reply that was fine.

Change any single category afterwards and the whole-side control reads Custom, which is a readout rather than a choice.

Choosing an action

When a reply is caught sets the action. The default is Replace unsafe replies.

ActionWhat happens to the reply
Watch onlyNothing. The reply is spoken exactly as it was written, and the hit is recorded
Replace unsafe repliesThe reply is cut and the agent speaks the safe line
Put the caller through to a personThe agent speaks the safe line, then transfers the caller
End the callThe agent speaks the safe line, then the call ends once the line finishes

Start on Watch only. It records what would have been caught and changes nothing the caller hears, so you can see how often a category fires on your own calls before it starts stopping replies.

A call ended by the End the call action is recorded as dropped, the same as an injection hang-up.

The safe line

Write your own line under Final safe response, up to 500 characters. Leave it empty and the agent uses the platform line:

Sorry, I can't help with that. Is there something else I can do for you?

The platform line is deliberately generic, because a substitute must not confirm what was caught. The line is spoken for every action except Watch only, and also when a check fails.

Putting the caller through to a person

This action speaks the safe line first, then hands the caller over. It rings the agent's Overflow number when one is set, otherwise the destination on its transfer action under Tools.

Without one, or on a channel with no phone leg such as an in-app test call, the agent speaks the safe line and the caller stays with the agent. The validator flags this action when no destination is set. If you write your own safe line for it, keep the line free of a promise to put the caller through, since the transfer is the part that can fail.

The whole reply is checked

Every clause of the reply is checked, not only the opening one, so a reply that starts politely and goes wrong three sentences later is still caught.

With any action other than Watch only, the reply is held back until its first clause has been checked, and each later piece is checked before it's released, so no unchecked words are spoken. The trade is time: the first word of audio waits on the first check, which is capped at two seconds. Under Watch only the reply plays straight away and is checked alongside.

A reply that ends before any clause boundary, such as a short unpunctuated answer, is checked as a whole before any of it is spoken.

What a caught reply sounds like

If an earlier clause is already playing, it stops the way it does when a caller interrupts. The rest of the reply is dropped, and the agent speaks the safe line once. Exactly one safe line is spoken per caught reply.

Topics that steer instead of stopping

Self-harm, child safety and regulated professional advice are switches rather than sensitivities, because there is no confidence score behind them to grade.

They're under Conversation guidance and apply on calls. Switching one on grades the recent conversation in the background every four turns and, where it finds drift, adds a short instruction that shapes the next reply. The caller waits on nothing, and the reply in flight is never interrupted.

Two consequences are worth planning for:

  • The correction arrives a turn late by design. If a topic has to stop a reply outright, it needs one of the five scored categories.
  • Grading runs an extra model call every four turns, which is why these three are off by default.

Business rules

To hold the agent to a rule about your business, such as never quoting a price or never naming a competitor, write the rule in the system prompt. Then add it to the Agent response guardrails card on this page as a question under Answer again when, such as "Does the reply quote a price?". Most replies follow the prompt. When one slips, the agent answers again with the rule in front of it, on calls and in chat. Agent response guardrails has the detail.

On this page