Glossary
Guardrail
A check that can allow, block, rewrite, or pause an agent action before it counts.
A guardrail is a check inside a harness. It runs on the way in, around a tool, or on the way out. The check can let the step through, block it, rewrite it, or pause it for a person.
OpenAI’s Agents SDK uses the word for automatic checks on input, output, or tool behavior. The same page treats human review as a separate pause: the run stops until a person or a policy approves or rejects a sensitive action. Pydantic AI’s harness docs describe a callable at the user prompt, at each tool call and result, and at the final output. The callable’s outcomes are allow, block, replace, retry, or — for a tool call — defer for approval.
How often a person must confirm a step is autonomy. The longer guide is What is an AI agent harness?.