In short
AI guardrails are controls that keep an AI system within limits: what it may say, which tools it may use and how much it may spend. The strongest guardrails are enforced by software outside the model, so the AI cannot talk its way past them.
Two kinds of guardrail
- Model-level: instructions, filters and output checks that shape what the AI says.
- System-level: permissions, spending caps and approvals enforced by the software around the AI.
Why system-level matters for agents
When an AI only talks, a bad answer is a bad answer. When it acts, a missed rule can cost money or reach customers. Actions need limits the model cannot override.
Common guardrail types
- Content filters on input and output.
- Tool allowlists: the only actions the agent may take.
- Spending limits per action and per month.
- Human approval for sensitive actions.
- Audit records of every step.
How to check a guardrail holds
Try to exceed it on purpose. A real limit refuses the action and records the refusal; a prompt-only rule may give way.
Guardrails in Maqron
Maqron checks balances, per-action limits, approvals and the monthly run allowance in its database before any cost is recorded.