Guardrails are controls around an AI system that limit unsafe, disallowed, or business-rule-breaking behavior.
They make safe behavior a system property, not only a sentence in a prompt.
Guardrails are controls used to limit unsafe, disallowed, or business-rule-breaking AI behavior.
They can inspect an input, restrict a tool call, enforce a permission, check an output, or stop an action before it reaches the outside world.
Guardrails can be placed around the model and around every action it proposes.
Inspect incoming requests before the model sees them.
Check generated content before it reaches a user or system.
Limit which tools an agent may call and how it may call them.
Define which data, systems, and actions are allowed.
Enforce rules such as amount limits, approval steps, or required fields.
Detect, block, or route unsafe or disallowed content.
A Coding Agent may be capable of many operations. Guardrails define which operations are actually available in this environment.
A prompt communicates a rule. A guardrail adds a control that can enforce the rule.
Review the plain-English vocabulary for Evaluation & Safety.
Open Evaluation & Safety Glossary →Guardrails turn safety rules into controls around inputs, outputs, tools, permissions, and business actions.