Compartilhar via


Usando ferramentas de função com aprovações humanas integradas

Esta etapa do tutorial mostra como usar ferramentas de funcionalidade que exigem aprovação humana com um agente, onde o agente é criado no Serviço de Conclusão de Bate-papo do Azure OpenAI.

Quando os agentes exigem qualquer entrada do usuário, por exemplo, para aprovar uma chamada de função, isso é conhecido como um padrão humano no loop. Uma execução de agente que requer entrada do usuário será concluída com uma resposta que indica qual entrada é necessária do usuário, em vez de concluir com uma resposta final. O chamador do agente é então responsável por obter a entrada necessária do usuário e passá-la de volta para o agente como parte de uma nova execução do agente.

Pré-requisitos

Para pré-requisitos e instalação de pacotes NuGet, consulte a etapa Criar e executar um agente simples neste tutorial.

Criar o agente com ferramentas funcionais

Ao usar funções, é possível indicar para cada função se ela requer aprovação humana antes de ser executada. Isso é feito encapsulando a AIFunction instância em uma ApprovalRequiredAIFunction instância.

Aqui está um exemplo de uma ferramenta de função simples que falsifica a obtenção do clima para um determinado local.

using System;
using System.ComponentModel;
using System.Linq;
using Azure.AI.OpenAI;
using Azure.Identity;
using Microsoft.Agents.AI;
using Microsoft.Extensions.AI;
using OpenAI;

[Description("Get the weather for a given location.")]
static string GetWeather([Description("The location to get the weather for.")] string location)
    => $"The weather in {location} is cloudy with a high of 15°C.";

Para criar um AIFunction e, em seguida, embrulhá-lo em um ApprovalRequiredAIFunction, você pode fazer o seguinte:

AIFunction weatherFunction = AIFunctionFactory.Create(GetWeather);
AIFunction approvalRequiredWeatherFunction = new ApprovalRequiredAIFunction(weatherFunction);

Ao criar o agente, agora você pode fornecer ao agente a ferramenta de função que exige aprovação, passando uma lista de ferramentas para o método AsAIAgent.

AIAgent agent = new AzureOpenAIClient(
    new Uri("https://<myresource>.openai.azure.com"),
    new AzureCliCredential())
     .GetChatClient("gpt-4o-mini")
     .AsAIAgent(instructions: "You are a helpful assistant", tools: [approvalRequiredWeatherFunction]);

Como agora você tem uma função que requer aprovação, o agente pode responder com uma solicitação de aprovação, em vez de executar a função diretamente e retornar o resultado. Você pode verificar o conteúdo da resposta para qualquer FunctionApprovalRequestContent instância, o que indica que o agente requer aprovação do usuário para uma função.

AgentSession session = await agent.CreateSessionAsync();
AgentResponse response = await agent.RunAsync("What is the weather like in Amsterdam?", session);

var functionApprovalRequests = response.Messages
    .SelectMany(x => x.Contents)
    .OfType<FunctionApprovalRequestContent>()
    .ToList();

Se houver solicitações de aprovação de função, os detalhes da chamada de função, incluindo nome e argumentos, poderão ser encontrados na FunctionCall propriedade na FunctionApprovalRequestContent instância. Isso pode ser mostrado ao usuário, para que ele possa decidir se aprova ou rejeita a chamada de função. Para este exemplo, suponha que haja uma solicitação.

FunctionApprovalRequestContent requestContent = functionApprovalRequests.First();
Console.WriteLine($"We require approval to execute '{requestContent.FunctionCall.Name}'");

Depois que o usuário tiver fornecido sua entrada, você poderá criar uma FunctionApprovalResponseContent instância usando o CreateResponse método no FunctionApprovalRequestContent. Passe true para aprovar a chamada de função ou false rejeitá-la.

O conteúdo da resposta pode então ser passado para o agente em um novo UserChatMessage, juntamente com o mesmo objeto de sessão para obter de volta do agente o resultado.

var approvalMessage = new ChatMessage(ChatRole.User, [requestContent.CreateResponse(true)]);
Console.WriteLine(await agent.RunAsync(approvalMessage, session));

Sempre que você estiver usando ferramentas de função com aprovações humanas no loop, lembre-se de verificar se há FunctionApprovalRequestContent instâncias na resposta, após cada execução do agente, até que todas as chamadas de função tenham sido aprovadas ou rejeitadas.

Dica

Consulte os exemplos do .NET para obter exemplos executáveis completos.

Esta etapa do tutorial mostra como usar ferramentas de função que exigem aprovação humana com um agente.

Quando os agentes exigem qualquer entrada do usuário, por exemplo, para aprovar uma chamada de função, isso é conhecido como um padrão humano no loop. Uma execução de agente que requer entrada do usuário será concluída com uma resposta que indica qual entrada é necessária do usuário, em vez de concluir com uma resposta final. O chamador do agente é então responsável por obter a entrada necessária do usuário e passá-la de volta para o agente como parte de uma nova execução do agente.

Pré-requisitos

Para pré-requisitos e instalação de pacotes do Python, consulte a etapa Criar e executar um agente simples neste tutorial.

Criar o agente com ferramentas funcionais que exigem aprovação

Ao usar funções, é possível indicar para cada função se ela requer aprovação humana antes de ser executada. Isso é feito definindo o parâmetro approval_mode para "always_require", ao usar o decorador @tool.

Aqui está um exemplo de uma ferramenta de função simples que falsifica a obtenção do clima para um determinado local.

from typing import Annotated
from agent_framework import tool

@tool
def get_weather(location: Annotated[str, "The city and state, e.g. San Francisco, CA"]) -> str:
    """Get the current weather for a given location."""
    return f"The weather in {location} is cloudy with a high of 15°C."

Para criar uma função que requer aprovação, você pode usar o approval_mode parâmetro:

@tool(approval_mode="always_require")
def get_weather_detail(location: Annotated[str, "The city and state, e.g. San Francisco, CA"]) -> str:
    """Get detailed weather information for a given location."""
    return f"The weather in {location} is cloudy with a high of 15°C, humidity 88%."

Ao criar o agente, agora você pode fornecer a ferramenta de função que requer aprovação para o agente, passando uma lista de ferramentas para o construtor Agent.

from agent_framework import Agent
from agent_framework.openai import OpenAIResponsesClient

async with Agent(
    chat_client=OpenAIResponsesClient(),
    name="WeatherAgent",
    instructions="You are a helpful weather assistant.",
    tools=[get_weather, get_weather_detail],
) as agent:
    # Agent is ready to use

Como agora você tem uma função que requer aprovação, o agente pode responder com uma solicitação de aprovação, em vez de executar a função diretamente e retornar o resultado. Você pode verificar a resposta de solicitações de entrada do usuário, o que indica que o agente requer aprovação do usuário para uma função.

result = await agent.run("What is the detailed weather like in Amsterdam?")

if result.user_input_requests:
    for user_input_needed in result.user_input_requests:
        print(f"Function: {user_input_needed.function_call.name}")
        print(f"Arguments: {user_input_needed.function_call.arguments}")

Se houver solicitações de aprovação de função, os detalhes da chamada de função, incluindo nome e argumentos, poderão ser encontrados na function_call propriedade na solicitação de entrada do usuário. Isso pode ser mostrado ao usuário, para que ele possa decidir se aprova ou rejeita a chamada de função.

Depois que o usuário tiver fornecido sua entrada, você poderá criar uma resposta usando o create_response método na solicitação de entrada do usuário. Passe True para aprovar a chamada de função ou False rejeitá-la.

Em seguida, a resposta pode ser passada em um novo Message para o agente, a fim de obter o resultado de volta.

from agent_framework import Message

# Get user approval (in a real application, this would be interactive)
user_approval = True  # or False to reject

# Create the approval response
approval_message = Message(
    role="user", 
    contents=[user_input_needed.create_response(user_approval)]
)

# Continue the conversation with the approval
final_result = await agent.run([
    "What is the detailed weather like in Amsterdam?",
    Message(role="assistant", contents=[user_input_needed]),
    approval_message
])
print(final_result.text)

Manipulando aprovações em um loop

Ao trabalhar com várias chamadas de função que exigem aprovação, talvez seja necessário lidar com aprovações em um loop até que todas as funções sejam aprovadas ou rejeitadas:

async def handle_approvals(query: str, agent) -> str:
    """Handle function call approvals in a loop."""
    current_input = query

    while True:
        result = await agent.run(current_input)

        if not result.user_input_requests:
            # No more approvals needed, return the final result
            return result.text

        # Build new input with all context
        new_inputs = [query]

        for user_input_needed in result.user_input_requests:
            print(f"Approval needed for: {user_input_needed.function_call.name}")
            print(f"Arguments: {user_input_needed.function_call.arguments}")

            # Add the assistant message with the approval request
            new_inputs.append(Message(role="assistant", contents=[user_input_needed]))

            # Get user approval (in practice, this would be interactive)
            user_approval = True  # Replace with actual user input

            # Add the user's approval response
            new_inputs.append(
                Message(role="user", contents=[user_input_needed.create_response(user_approval)])
            )

        # Continue with all the context
        current_input = new_inputs

# Usage
result_text = await handle_approvals("Get detailed weather for Seattle and Portland", agent)
print(result_text)

Sempre que você estiver usando ferramentas de função com aprovações humanas no loop, lembre-se de verificar se há solicitações de entrada do usuário na resposta, após cada execução do agente, até que todas as chamadas de função tenham sido aprovadas ou rejeitadas.

Exemplo completo

# Copyright (c) Microsoft. All rights reserved.

import asyncio
from random import randrange
from typing import TYPE_CHECKING, Annotated, Any

from agent_framework import Agent, AgentResponse, Message, tool
from agent_framework.openai import OpenAIResponsesClient

if TYPE_CHECKING:
    from agent_framework import SupportsAgentRun

"""
Demonstration of a tool with approvals.

This sample demonstrates using AI functions with user approval workflows.
It shows how to handle function call approvals without using threads.
"""

conditions = ["sunny", "cloudy", "raining", "snowing", "clear"]


# NOTE: approval_mode="never_require" is for sample brevity. Use "always_require" in production; see samples/02-agents/tools/function_tool_with_approval.py and samples/02-agents/tools/function_tool_with_approval_and_sessions.py.
@tool(approval_mode="never_require")
def get_weather(location: Annotated[str, "The city and state, e.g. San Francisco, CA"]) -> str:
    """Get the current weather for a given location."""
    # Simulate weather data
    return f"The weather in {location} is {conditions[randrange(0, len(conditions))]} and {randrange(-10, 30)}°C."


# Define a simple weather tool that requires approval
@tool(approval_mode="always_require")
def get_weather_detail(location: Annotated[str, "The city and state, e.g. San Francisco, CA"]) -> str:
    """Get the current weather for a given location."""
    # Simulate weather data
    return (
        f"The weather in {location} is {conditions[randrange(0, len(conditions))]} and {randrange(-10, 30)}°C, "
        "with a humidity of 88%. "
        f"Tomorrow will be {conditions[randrange(0, len(conditions))]} with a high of {randrange(-10, 30)}°C."
    )


async def handle_approvals(query: str, agent: "SupportsAgentRun") -> AgentResponse:
    """Handle function call approvals.

    When we don't have a thread, we need to ensure we include the original query,
    the approval request, and the approval response in each iteration.
    """
    result = await agent.run(query)
    while len(result.user_input_requests) > 0:
        # Start with the original query
        new_inputs: list[Any] = [query]

        for user_input_needed in result.user_input_requests:
            print(
                f"\nUser Input Request for function from {agent.name}:"
                f"\n  Function: {user_input_needed.function_call.name}"
                f"\n  Arguments: {user_input_needed.function_call.arguments}"
            )

            # Add the assistant message with the approval request
            new_inputs.append(Message("assistant", [user_input_needed]))

            # Get user approval
            user_approval = await asyncio.to_thread(input, "\nApprove function call? (y/n): ")

            # Add the user's approval response
            new_inputs.append(
                Message("user", [user_input_needed.to_function_approval_response(user_approval.lower() == "y")])
            )

        # Run again with all the context
        result = await agent.run(new_inputs)

    return result


async def handle_approvals_streaming(query: str, agent: "SupportsAgentRun") -> None:
    """Handle function call approvals with streaming responses.

    When we don't have a thread, we need to ensure we include the original query,
    the approval request, and the approval response in each iteration.
    """
    current_input: str | list[Any] = query
    has_user_input_requests = True
    while has_user_input_requests:
        has_user_input_requests = False
        user_input_requests: list[Any] = []

        # Stream the response
        async for chunk in agent.run(current_input, stream=True):
            if chunk.text:
                print(chunk.text, end="", flush=True)

            # Collect user input requests from the stream
            if chunk.user_input_requests:
                user_input_requests.extend(chunk.user_input_requests)

        if user_input_requests:
            has_user_input_requests = True
            # Start with the original query
            new_inputs: list[Any] = [query]

            for user_input_needed in user_input_requests:
                print(
                    f"\n\nUser Input Request for function from {agent.name}:"
                    f"\n  Function: {user_input_needed.function_call.name}"
                    f"\n  Arguments: {user_input_needed.function_call.arguments}"
                )

                # Add the assistant message with the approval request
                new_inputs.append(Message("assistant", [user_input_needed]))

                # Get user approval
                user_approval = await asyncio.to_thread(input, "\nApprove function call? (y/n): ")

                # Add the user's approval response
                new_inputs.append(
                    Message("user", [user_input_needed.to_function_approval_response(user_approval.lower() == "y")])
                )

            # Update input with all the context for next iteration
            current_input = new_inputs


async def run_weather_agent_with_approval(stream: bool) -> None:
    """Example showing AI function with approval requirement."""
    print(f"\n=== Weather Agent with Approval Required ({'Streaming' if stream else 'Non-Streaming'}) ===\n")

    async with Agent(
        client=OpenAIResponsesClient(),
        name="WeatherAgent",
        instructions=("You are a helpful weather assistant. Use the get_weather tool to provide weather information."),
        tools=[get_weather, get_weather_detail],
    ) as agent:
        query = "Can you give me an update of the weather in LA and Portland and detailed weather for Seattle?"
        print(f"User: {query}")

        if stream:
            print(f"\n{agent.name}: ", end="", flush=True)
            await handle_approvals_streaming(query, agent)
            print()
        else:
            result = await handle_approvals(query, agent)
            print(f"\n{agent.name}: {result}\n")


async def main() -> None:
    print("=== Demonstration of a tool with approvals ===\n")

    await run_weather_agent_with_approval(stream=False)
    await run_weather_agent_with_approval(stream=True)


if __name__ == "__main__":
    asyncio.run(main())
# Copyright (c) Microsoft. All rights reserved.

import asyncio
from random import randrange
from typing import TYPE_CHECKING, Annotated, Any

from agent_framework import Agent, AgentResponse, Message, tool
from agent_framework.openai import OpenAIResponsesClient

if TYPE_CHECKING:
    from agent_framework import SupportsAgentRun

"""
Demonstration of a tool with approvals.

This sample demonstrates using AI functions with user approval workflows.
It shows how to handle function call approvals without using threads.
"""

conditions = ["sunny", "cloudy", "raining", "snowing", "clear"]


# NOTE: approval_mode="never_require" is for sample brevity. Use "always_require" in production; see samples/02-agents/tools/function_tool_with_approval.py and samples/02-agents/tools/function_tool_with_approval_and_sessions.py.
@tool(approval_mode="never_require")
def get_weather(location: Annotated[str, "The city and state, e.g. San Francisco, CA"]) -> str:
    """Get the current weather for a given location."""
    # Simulate weather data
    return f"The weather in {location} is {conditions[randrange(0, len(conditions))]} and {randrange(-10, 30)}°C."


# Define a simple weather tool that requires approval
@tool(approval_mode="always_require")
def get_weather_detail(location: Annotated[str, "The city and state, e.g. San Francisco, CA"]) -> str:
    """Get the current weather for a given location."""
    # Simulate weather data
    return (
        f"The weather in {location} is {conditions[randrange(0, len(conditions))]} and {randrange(-10, 30)}°C, "
        "with a humidity of 88%. "
        f"Tomorrow will be {conditions[randrange(0, len(conditions))]} with a high of {randrange(-10, 30)}°C."
    )


async def handle_approvals(query: str, agent: "SupportsAgentRun") -> AgentResponse:
    """Handle function call approvals.

    When we don't have a thread, we need to ensure we include the original query,
    the approval request, and the approval response in each iteration.
    """
    result = await agent.run(query)
    while len(result.user_input_requests) > 0:
        # Start with the original query
        new_inputs: list[Any] = [query]

        for user_input_needed in result.user_input_requests:
            print(
                f"\nUser Input Request for function from {agent.name}:"
                f"\n  Function: {user_input_needed.function_call.name}"
                f"\n  Arguments: {user_input_needed.function_call.arguments}"
            )

            # Add the assistant message with the approval request
            new_inputs.append(Message("assistant", [user_input_needed]))

            # Get user approval
            user_approval = await asyncio.to_thread(input, "\nApprove function call? (y/n): ")

            # Add the user's approval response
            new_inputs.append(
                Message("user", [user_input_needed.to_function_approval_response(user_approval.lower() == "y")])
            )

        # Run again with all the context
        result = await agent.run(new_inputs)

    return result


async def handle_approvals_streaming(query: str, agent: "SupportsAgentRun") -> None:
    """Handle function call approvals with streaming responses.

    When we don't have a thread, we need to ensure we include the original query,
    the approval request, and the approval response in each iteration.
    """
    current_input: str | list[Any] = query
    has_user_input_requests = True
    while has_user_input_requests:
        has_user_input_requests = False
        user_input_requests: list[Any] = []

        # Stream the response
        async for chunk in agent.run(current_input, stream=True):
            if chunk.text:
                print(chunk.text, end="", flush=True)

            # Collect user input requests from the stream
            if chunk.user_input_requests:
                user_input_requests.extend(chunk.user_input_requests)

        if user_input_requests:
            has_user_input_requests = True
            # Start with the original query
            new_inputs: list[Any] = [query]

            for user_input_needed in user_input_requests:
                print(
                    f"\n\nUser Input Request for function from {agent.name}:"
                    f"\n  Function: {user_input_needed.function_call.name}"
                    f"\n  Arguments: {user_input_needed.function_call.arguments}"
                )

                # Add the assistant message with the approval request
                new_inputs.append(Message("assistant", [user_input_needed]))

                # Get user approval
                user_approval = await asyncio.to_thread(input, "\nApprove function call? (y/n): ")

                # Add the user's approval response
                new_inputs.append(
                    Message("user", [user_input_needed.to_function_approval_response(user_approval.lower() == "y")])
                )

            # Update input with all the context for next iteration
            current_input = new_inputs


async def run_weather_agent_with_approval(stream: bool) -> None:
    """Example showing AI function with approval requirement."""
    print(f"\n=== Weather Agent with Approval Required ({'Streaming' if stream else 'Non-Streaming'}) ===\n")

    async with Agent(
        client=OpenAIResponsesClient(),
        name="WeatherAgent",
        instructions=("You are a helpful weather assistant. Use the get_weather tool to provide weather information."),
        tools=[get_weather, get_weather_detail],
    ) as agent:
        query = "Can you give me an update of the weather in LA and Portland and detailed weather for Seattle?"
        print(f"User: {query}")

        if stream:
            print(f"\n{agent.name}: ", end="", flush=True)
            await handle_approvals_streaming(query, agent)
            print()
        else:
            result = await handle_approvals(query, agent)
            print(f"\n{agent.name}: {result}\n")


async def main() -> None:
    print("=== Demonstration of a tool with approvals ===\n")

    await run_weather_agent_with_approval(stream=False)
    await run_weather_agent_with_approval(stream=True)


if __name__ == "__main__":
    asyncio.run(main())

Próximas etapas