input_guardrail is skipped

[Salaudev]

@input_guardrail
async def rate_info_guardrail(
    ctx: RunContextWrapper[OrchestratorContext],  # your context type
    agent: Agent,
    input: str | list[TResponseInputItem],  # same sig the SDK expects
) -> GuardrailFunctionOutput:
    """
    Abort the run if any critical rate-info fields are missing, or if
    is_bid_request_sent is still False.
    """

    _REQUIRED_RATE_FIELDS = [
        "maximum_rate",
        "minimum_rate",
        "rate_usd",
        "is_bid_request_sent",
    ]

    rate = ctx.context.load_context.rate_info  # <-- your own object

    # 2️⃣  Collect the names of missing / invalid fields.
    missing: list[str] = []
    for name in _REQUIRED_RATE_FIELDS:
        value = getattr(rate, name, None)
        if value is None:  # not set
            missing.append(name)
        elif name == "is_bid_request_sent" and value is False:
            missing.append(name)

    # 3️⃣  Decide whether to trip the wire.
    trip = bool(missing)

    print("Rate info guardrail triggered...")

    # 4️⃣  Return the standard GuardrailFunctionOutput.
    return GuardrailFunctionOutput(
        output_info={"missing_fields": missing},  # great for debugging/logging
        tripwire_triggered=trip,
    )


negotiation_agent = Agent[OrchestratorContext](
    name="Negotiation Agent",
    instructions=dynamic_negotiation_agent_instructions,
    model=OpenAIChatCompletionsModel(
        model=os.getenv("AZURE_OPENAI_DEPLOYMENT_NAME"), openai_client=azure_client
    ),
    tools=[
        FunctionTool(
            name="send_reply",
            on_invoke_tool=send_reply,
            description="This tool will send the HTML email body to the broker",
            params_json_schema=AgentsReq.model_json_schema(),
        ),
    ],
    input_guardrails=[rate_info_guardrail],
)

In my logs I see that my negotiation_agent is running, but I don't see the print("Rate info guardrail triggered...")

[rm-openai]

@Salaudev is the negotiation_agent the first agent in the run? Only then is the input guadrail called

[Salaudev]

okay, thanks @rm-openai

now I got another issue

i am going to implement it using manager based way, that mean calling agent as a tool

This is my manager agent

# should orchestrator between agents
manager = Agent[OrchestratorContext](
    name="Manager Agent",
    instructions="""
        ## Role ##
        You are a professional dispatcher in a trucking company.
        You are responsible for orchestrating the entire load management
        process from analysis to negotiation and final decision-making.

        Your tasks:
        - Use the extract_details tool to gather all load details and to check missing fields from conversations and what information is missing
        - Use the analyze_warnings tool to identify any compliance issues or concerns
        - Use the answer_question tool to answer any questions the broker may have
        - Ask the broker if we have any missed details
        - Use the communicate_with_broker tool to communicate with the broker
        - Use the cancel_load tool to cancel the load
    """,
    model=OpenAIChatCompletionsModel(
        model="gpt-4o", openai_client=azure_client
    ),
    tools=[
        FunctionTool(
            name="cancel_load",
            on_invoke_tool=cancel_load,
            description="This tool will cancel the load",
            params_json_schema=AgentsReq.model_json_schema(),
        ),
        extractor.as_tool(
            tool_name="extract_details",
            tool_description="This tool will extract the load details from the conversation history",
        ),
        warnings_analyzer.as_tool(
            tool_name="analyze_warnings",
            tool_description="This tool will analyze the load warnings",
        ),
        question_answer.as_tool(
            tool_name="answer_question",
            tool_description="This tool will find the answer to the broker's question using the provided load details and company information",
        ),
        communication_agent.as_tool(
            tool_name="communicate_with_broker",
            tool_description="This tool will send a message to the broker",
        ),
    ],
)

In the logs I see this

We started the conversation and we added this message to the conversation history with role assistant

And when we get reply , agent is runnings and
gives the prev message to into child agents, why it's working like that. how can I avoid that ?

[adhishthite]

@Salaudev Curious, what tool are you using for viz?

[Salaudev]

@Salaudev Curious, what tool are you using for viz?

Pydantic logfire

[Salaudev]

@rm-openai can you help with this please

[rm-openai]

@Salaudev by default, agent.as_tool has a single input: str that the calling agent has to fill in. You can either give it instructions on what to send (e.g. in the tool description), or create a custom agent-as-tool like this:

https://openai.github.io/openai-agents-python/tools/#customizing-tool-agents