Sdílet prostřednictvím


Pracovní postupy rozhraní Microsoft Agent Framework – Práce s agenty

Tato stránka obsahuje přehled použití agentů v rámci pracovních postupů rozhraní Microsoft Agent Framework.

Přehled

K přidání inteligentních informací do pracovních postupů můžete využít agenty AI jako součást provádění pracovního postupu. Agenty umělé inteligence je možné snadno integrovat do pracovních postupů, což umožňuje vytvářet složitá a inteligentní řešení, která byla dříve obtížně dosažitelná.

Přidání agenta přímo do pracovního postupu

Agenty můžete do pracovního postupu přidat přes hrany:

using Microsoft.Agents.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>();

Spuštění pracovního postupu

Uvnitř výše vytvořeného pracovního postupu jsou agenti ve skutečnosti zabaleni uvnitř exekutoru, který zpracovává komunikaci agenta s jinými částmi pracovního postupu. Exekutor dokáže zpracovat tři typy zpráv:

  • ChatMessage: Jedna chatovací zpráva
  • List<ChatMessage>: Seznam zpráv chatu
  • TurnToken: Turn token, který signalizuje začátek nového turnu

Exekutor neaktivuje agenta, aby reagoval, dokud neobdrží TurnToken. Jakékoliv zprávy přijaté před TurnToken jsou vyrovnávány a odeslány agentovi, když je přijato 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}");
    }
}

Použití integrovaného exekutoru agenta

Agenty můžete do pracovního postupu přidat přes hrany:

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()

Spuštění pracovního postupu

Uvnitř výše vytvořeného pracovního postupu jsou agenti ve skutečnosti zabaleni uvnitř exekutoru, který zpracovává komunikaci agenta s jinými částmi pracovního postupu. Exekutor dokáže zpracovat tři typy zpráv:

  • str: Jedna chatovací zpráva ve formátu řetězce
  • ChatMessage: Jedna chatovací zpráva
  • List<ChatMessage>: Seznam zpráv chatu

Kdykoli exekutor obdrží zprávu jednoho z těchto typů, aktivuje agenta, aby odpověděl, a typ odpovědi bude AgentExecutorResponse objektem. Tato třída obsahuje užitečné informace o odpovědi agenta, včetně:

  • executor_id: ID exekutoru, který vytvořil tuto odpověď.
  • agent_run_response: Úplná odpověď od agenta
  • full_conversation: Úplná historie konverzací až do tohoto okamžiku

Při spuštění pracovního postupu je možné vygenerovat dva možné typy událostí související s odpověďmi agentů:

  • AgentRunUpdateEvent obsahující bloky odpovědí agenta, které se generují v režimu streamování.
  • AgentRunEvent obsahující úplnou odpověď agenta v režimu bez streamování.

Ve výchozím nastavení jsou agenti zabaleni do exekutorů, které běží v režimu streamování. Toto chování můžete přizpůsobit vytvořením vlastního exekutoru. Další podrobnosti najdete v další části.

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)

Použití vlastního spouštěče agenta

Někdy můžete chtít přizpůsobit, jak jsou agenti AI integrováni do pracovního postupu. Toho dosáhnete vytvořením vlastního exekutoru. To vám umožní řídit:

  • Vyvolání agenta: streamování nebo ne-streamování
  • Typy zpráv, které bude agent zpracovávat, včetně vlastních typů zpráv
  • Životní cyklus agenta, včetně inicializace a čištění
  • Použití agentních vláken a dalších prostředků
  • Další události generované během běhu agenta, včetně uživatelských událostí
  • Integrace s jinými funkcemi pracovního postupu, jako jsou sdílené stavy a požadavky/odpovědi
internal sealed class CustomAgentExecutor : ReflectingExecutor<CustomAgentExecutor>, IMessageHandler<CustomInput, CustomOutput>
{
    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)

Další kroky