Guardrails
Built-in protection that keeps every agent inside its defined purpose — automatic, always on, nothing to configure.
Every Talkr agent ships with a baseline set of protections against callers trying to talk it off-script. This is automatic — it's not a toggle, a setting, or something you can accidentally leave off. Every agent has it, on every call.
What it blocks
Revealing internal details
Your agent won't share, quote, or discuss its system prompt, instructions, internal configuration, model, provider, or settings like temperature — even if a caller asks directly or insists it already told them. It responds that it can't share internal details and redirects to how it can help.
Claiming to change its own behavior
A caller can't talk your agent into "changing its temperature," "switching models," or otherwise reconfiguring itself — because it can't actually do that, and it won't pretend to.
Going off-script
If a caller tries to get the agent to ignore its instructions, roleplay as something else, or answer something unrelated to why it's on the call, it politely declines and steers back to its actual purpose.
Knowledge Base grounding is also part of this — see Knowledge Base: Reducing Hallucinations for how your agent avoids guessing at facts it can't confirm from your documents.
How it works
This is a baseline instruction built into every agent's behavior, not something you write into a node prompt yourself — you don't need to add "don't reveal your system prompt" language to your own prompts; it's already covered. Your own prompt still fully controls what the agent actually does and says on-topic — this only holds the line on what's out of bounds.
Catching what gets through
No prompt-level protection is airtight against every possible attempt. As a second layer, QA can automatically flag calls after the fact if something did leak through — an agent revealing internal details, drifting off-topic, or stating a fact it couldn't actually back up. Turn on Alert on guardrail violation under Settings → QA Alerts to get notified when that happens, the same way QA already alerts on low scores or negative sentiment.