Glossary
Guardrails
Guardrails are the constraints placed around an AI agent — what it may do, what it must ask about, what it can spend and what is recorded — so its behaviour stays within bounds you chose.
The four that matter
- Capability limits — which tools it holds at all. The strongest guardrail is a tool the agent does not have.
- Approval gates — which actions wait for a person.
- Spending caps — hard rather than advisory, because an agent on a schedule can run up a bill.
- Audit — a record of what happened, without which none of the others can be verified.
Not the same as a system prompt
Telling an agent not to do something is guidance, not a guardrail. A guardrail is enforced outside the model, so it holds regardless of what the model decides. If the only thing stopping an action is an instruction, there is no guardrail.
Proportionate, not maximal
Guardrails tight enough to stop useful work get disabled. The aim is matching the constraint to the consequence.
Related terms
FAQ
What are AI guardrails?
The constraints placed around an agent — which tools it holds, which actions wait for human approval, what it may spend, and what is recorded — so behaviour stays within bounds you chose rather than bounds the model chose.
Is a system prompt a guardrail?
No. Telling an agent not to do something is guidance. A guardrail is enforced outside the model, so it holds regardless of what the model decides. If the only thing preventing an action is an instruction, there is no guardrail.