What are guardrails?

The word guardrails is used loosely across vendors and has no single specification. This entry follows the OpenAI Agents SDK documentation, which is one vendor's description of the idea. Other frameworks use the term for related but different mechanisms.

In that SDK, guardrails are checks and validations on what goes into an agent and what comes out. The docs give the example of a fast, cheap model screening a request before a slower, expensive agent runs. The docs describe input and output guardrails, and add tool guardrails for function-tool calls. Input guardrails check the initial user input and run only for the first agent in a chain. Output guardrails check the final output and run only for the agent that produces it. Tool guardrails wrap function tools and run on each call, before or after execution.

When a check fails it signals a tripwire. The runner then raises an exception and halts execution. Input guardrails run in parallel with the agent by default, which is faster but means the agent may already have used tokens or run tools before the tripwire fires. A blocking mode finishes the check first.

The page lists limits: tool guardrails do not apply to handoff calls, hosted tools or built-in execution tools, and it says Agent.as_tool() does not currently expose tool-guardrail options directly.