Guardrails in Agent Canvas: Structural Controls for AI Agents
Guardrails in Agent Canvas aren't prompt instructions — they're hard constraints in the conversation graph that the LLM cannot bypass. Here's how they work.
When you tell an LLM 'never provide medical advice,' you're relying on probabilistic compliance. When you add a guardrail node in Agent Canvas that detects medical topics and routes to an escalation flow, you're enforcing a structural constraint. The difference: one occasionally fails. The other cannot be bypassed by the model.
Types of guardrails
- Topic blocks — prevent the agent from engaging with specific subjects (competitor pricing, medical diagnosis, legal advice). Detection happens at the orchestration layer, not the prompt.
- Mandatory disclosures — inject required language at specific points: AI identity disclosure at call start, compliance language before financial transactions, recording consent notifications.
- Escalation triggers — automatic human handoff when conditions are met: customer requests human, sentiment drops below threshold, conversation exceeds complexity limit.
- Data handling rules — redact sensitive information from logs, prevent storage of payment data, enforce field-level encryption for PII.
- Action limits — restrict which tools the agent can use and under what conditions (e.g., can process refunds under $50 autonomously, must escalate above $50).
Guardrails are visible in the canvas
Every guardrail is a visible node or edge in the conversation graph. Compliance teams can audit the flow by inspecting the canvas — they can see where disclosures are injected, where escalation happens, and what topic filters are active. This auditability is impossible with prompt-only systems where guardrails are buried in paragraphs of natural language instructions that may or may not be followed.
Ready to build?
See how Mazed's multimodal AI agents work for your use case.