Поделиться через


Рабочие процессы Microsoft Agent Framework — работа с агентами

На этой странице представлен обзор использования агентов в рабочих процессах Microsoft Agent Framework.

Обзор

Чтобы добавить аналитику в рабочие процессы, вы можете использовать агенты ИИ в рамках выполнения рабочего процесса. Агенты искусственного интеллекта можно легко интегрировать в рабочие процессы, что позволяет создавать сложные интеллектуальные решения, которые ранее были трудно достичь.

Добавление агента непосредственно в рабочий процесс

Агенты можно добавить в рабочий процесс через узлы.

using Microsoft.Agents.AI.Workflows;
using Microsoft.Extensions.AI;
using Microsoft.Agents.AI;

// Create the agents first
AIAgent agentA = new ChatClientAgent(chatClient, instructions);
AIAgent agentB = new ChatClientAgent(chatClient, instructions);

// Build a workflow with the agents
WorkflowBuilder builder = new(agentA);
builder.AddEdge(agentA, agentB);
Workflow<ChatMessage> workflow = builder.Build<ChatMessage>();

Запуск рабочего процесса

В рабочем процессе, созданном выше, агенты фактически обернуты внутри исполнителя, который обрабатывает их коммуникацию с другими частями рабочего процесса. Исполнитель может обрабатывать три типа сообщений:

  • ChatMessage: одно сообщение чата
  • List<ChatMessage>: список сообщений чата
  • TurnToken: маркер поворота, который сигнализирует о начале нового поворота

Исполнитель не активирует агент, пока он не получит ответ TurnToken. Все сообщения, полученные до TurnToken, буферизуются и отправляются агенту, когда получено TurnToken.

StreamingRun run = await InProcessExecution.StreamAsync(workflow, new ChatMessage(ChatRole.User, "Hello World!"));
// Must send the turn token to trigger the agents. The agents are wrapped as executors.
// When they receive messages, they will cache the messages and only start processing
// when they receive a TurnToken. The turn token will be passed from one agent to the next.
await run.TrySendMessageAsync(new TurnToken(emitEvents: true));
await foreach (WorkflowEvent evt in run.WatchStreamAsync().ConfigureAwait(false))
{
    // The agents will run in streaming mode and an AgentRunUpdateEvent
    // will be emitted as new chunks are generated.
    if (evt is AgentRunUpdateEvent agentRunUpdate)
    {
        Console.WriteLine($"{agentRunUpdate.ExecutorId}: {agentRunUpdate.Data}");
    }
}

Использование встроенного исполнителя агента

Агенты можно добавить в рабочий процесс через узлы.

from agent_framework import WorkflowBuilder
from agent_framework.azure import AzureChatClient
from azure.identity import AzureCliCredential

# Create the agents first
chat_client = AzureChatClient(credential=AzureCliCredential())
writer_agent: ChatAgent = chat_client.create_agent(
    instructions=(
        "You are an excellent content writer. You create new content and edit contents based on the feedback."
    ),
    name="writer_agent",
)
reviewer_agent = chat_client.create_agent(
    instructions=(
        "You are an excellent content reviewer."
        "Provide actionable feedback to the writer about the provided content."
        "Provide the feedback in the most concise manner possible."
    ),
    name="reviewer_agent",
)

# Build a workflow with the agents
builder = WorkflowBuilder()
builder.set_start_executor(writer_agent)
builder.add_edge(writer_agent, reviewer_agent)
workflow = builder.build()

Запуск рабочего процесса

В рабочем процессе, созданном выше, агенты фактически обернуты внутри исполнителя, который обрабатывает их коммуникацию с другими частями рабочего процесса. Исполнитель может обрабатывать три типа сообщений:

  • str: одно сообщение чата в строковом формате
  • ChatMessage: одно сообщение чата
  • List<ChatMessage>: список сообщений чата

Каждый раз, когда исполнитель получает сообщение одного из этих типов, он заставляет агента реагировать, а тип ответа будет AgentExecutorResponse объектом. Этот класс содержит полезные сведения об ответе агента, в том числе:

  • executor_id: идентификатор исполнителя, создающего этот ответ
  • agent_run_response: полный ответ от агента
  • full_conversation: вся история бесед до этой точки

При запуске рабочего процесса можно создать два возможных типа событий, связанных с ответами агентов:

  • AgentRunUpdateEvent содержит фрагменты ответа агента по мере их создания в режиме потоковой передачи.
  • AgentRunEvent содержит полный ответ от агента в режиме, отличном от потоковой передачи.

По умолчанию агенты упаковываются в исполнителей, которые выполняются в режиме потоковой передачи. Это поведение можно настроить, создав пользовательский исполнительный механизм. Дополнительные сведения см. в следующем разделе.

last_executor_id = None
async for event in workflow.run_streaming("Write a short blog post about AI agents."):
    if isinstance(event, AgentRunUpdateEvent):
        if event.executor_id != last_executor_id:
            if last_executor_id is not None:
                print()
            print(f"{event.executor_id}:", end=" ", flush=True)
            last_executor_id = event.executor_id
        print(event.data, end="", flush=True)

Использование пользовательского исполнителя агента

Иногда может потребоваться настроить интеграцию агентов ИИ в рабочий процесс. Это можно сделать, создав пользовательский выполнитель. Это позволяет управлять следующими способами:

  • Вызов агента: потоковый или непотоковый режим
  • Типы сообщений, которые агент будет обрабатывать, включая пользовательские типы сообщений
  • Жизненный цикл агента, включая инициализацию и очистку
  • Использование потоков агентов и других ресурсов
  • Дополнительные события, создаваемые во время выполнения агента, включая пользовательские события
  • Интеграция с другими функциями рабочего процесса, такими как общие состояния и запросы и ответы
internal sealed class CustomAgentExecutor : Executor<CustomInput, CustomOutput>("CustomAgentExecutor")
{
    private readonly AIAgent _agent;

    /// <summary>
    /// Creates a new instance of the <see cref="CustomAgentExecutor"/> class.
    /// </summary>
    /// <param name="agent">The AI agent used for custom processing</param>
    public CustomAgentExecutor(AIAgent agent) : base("CustomAgentExecutor")
    {
        this._agent = agent;
    }

    public async ValueTask<CustomOutput> HandleAsync(CustomInput message, IWorkflowContext context)
    {
        // Retrieve any shared states if needed
        var sharedState = await context.ReadStateAsync<SharedStateType>("sharedStateId", scopeName: "SharedStateScope");

        // Render the input for the agent
        var agentInput = RenderInput(message, sharedState);

        // Invoke the agent
        // Assume the agent is configured with structured outputs with type `CustomOutput`
        var response = await this._agent.RunAsync(agentInput);
        var customOutput = JsonSerializer.Deserialize<CustomOutput>(response.Text);

        return customOutput;
    }
}
from agent_framework import (
    ChatAgent,
    ChatMessage,
    Executor,
    WorkflowContext,
    handler
)

class Writer(Executor):

    agent: ChatAgent

    def __init__(self, chat_client: AzureChatClient, id: str = "writer"):
        # Create a domain specific agent using your configured AzureChatClient.
        agent = chat_client.create_agent(
            instructions=(
                "You are an excellent content writer. You create new content and edit contents based on the feedback."
            ),
        )
        # Associate the agent with this executor node. The base Executor stores it on self.agent.
        super().__init__(agent=agent, id=id)

    @handler
    async def handle(self, message: ChatMessage, ctx: WorkflowContext[list[ChatMessage]]) -> None:
        """Handles a single chat message and forwards the accumulated messages to the next executor in the workflow."""
        # Invoke the agent with the incoming message and get the response
        messages: list[ChatMessage] = [message]
        response = await self.agent.run(messages)
        # Accumulate messages and send them to the next executor in the workflow.
        messages.extend(response.messages)
        await ctx.send_message(messages)

Дальнейшие шаги