What happens when an OpenAI Agents SDK guardrail trips?
An agent-level tripwire stops the run and raises InputGuardrailTripwireTriggered or OutputGuardrailTripwireTriggered. A blocking input guardrail catches the request before the agent starts; the default parallel mode can spend tokens or call a tool before the tripwire lands. A function-tool guardrail has another choice: reject the call or result and let the agent continue, or raise a tool-scoped tripwire exception and stop it. The SDK documents both paths.
The decision survives the exception. An agent tripwire carries guardrail_result, and its input or output results include the check that fired. A tool tripwire carries guardrail and output directly; its run_data lists hold earlier completed tool checks, not the triggering result. The SDK documents the exception fields. Counting trips from the agent-level result list alone will miss tool trips.