Input guardrail agent

input-guardrail-agent

Decorate a function with @input_guardrail and attach it to an agent's input_guardrails list. If the guardrail's tripwire_triggered=True, an InputGuardrailTripwireTriggered exception is raised and the main agent never runs.

When to apply

Cost-control (block expensive model before it runs), content moderation (off-topic check), or compliance (detect PII/malicious input). Run fast/cheap checks before the expensive main agent.

Core snippet

from pydantic import BaseModel
from agents import (
    Agent, GuardrailFunctionOutput, InputGuardrailTripwireTriggered,
    RunContextWrapper, Runner, TResponseInputItem, input_guardrail,
)

class MathHomeworkOutput(BaseModel):
    reasoning: str
    is_math_homework: bool

guardrail_agent = Agent(
    name="Guardrail check",
    instructions="Check if the user is asking you to do their math homework.",
    output_type=MathHomeworkOutput,
)

@input_guardrail
async def math_guardrail(
    context: RunContextWrapper[None], agent: Agent, input: str | list[TResponseInputItem]
) -> GuardrailFunctionOutput:
    result = await Runner.run(guardrail_agent, input, context=context.context)
    final_output = result.final_output_as(MathHomeworkOutput)
    return GuardrailFunctionOutput(
        output_info=final_output,
        tripwire_triggered=final_output.is_math_homework,
    )

main_agent = Agent(
    name="Customer support agent",
    instructions="You are a customer support agent.",
    input_guardrails=[math_guardrail],
)

async def main():
    try:
        result = await Runner.run(main_agent, "Can you help me solve 2+2?")
        print(result.final_output)
    except InputGuardrailTripwireTriggered:
        print("Guardrail triggered! This looks like math homework.")

Key notes

• Input guardrails run only when the agent is the first agent in the chain
• Default mode: runs in parallel with the agent for best latency
• For blocking mode (agent never starts if guardrail fails): @input_guardrail(run_in_parallel=False)
• The guardrail function can be async or sync
• output_info is stored in the exception for downstream handling