Dauerhafte Aufgabenerweiterung für Microsoft Agent Framework (Vorschau)

var endpoint = Environment.GetEnvironmentVariable("AZURE_OPENAI_ENDPOINT")
    ?? throw new InvalidOperationException("AZURE_OPENAI_ENDPOINT is not set.");
var deploymentName = Environment.GetEnvironmentVariable("AZURE_OPENAI_DEPLOYMENT_NAME")
    ?? throw new InvalidOperationException("AZURE_OPENAI_DEPLOYMENT_NAME is not set.");

AIAgent agent = new AzureOpenAIClient(new Uri(endpoint), new DefaultAzureCredential())
    .GetChatClient(deploymentName)
    .AsAIAgent(
        instructions: "You are a professional content writer who creates engaging, "
                    + "well-structured documents for any given topic.",
        name: "DocumentPublisher");

// One line to make the agent durable with serverless hosting
using IHost app = FunctionsApplication
    .CreateBuilder(args)
    .ConfigureFunctionsWebApplication()
    .ConfigureDurableAgents(options => options.AddAIAgent(agent))
    .Build();
app.Run();

import os
from agent_framework.azure import AzureOpenAIChatClient, AgentFunctionApp
from azure.identity import DefaultAzureCredential

endpoint = os.getenv("AZURE_OPENAI_ENDPOINT")
deployment_name = os.getenv("AZURE_OPENAI_DEPLOYMENT_NAME")

agent = AzureOpenAIChatClient(
    endpoint=endpoint,
    deployment_name=deployment_name,
    credential=DefaultAzureCredential()
).as_agent(
    instructions="You are a professional content writer who creates engaging, "
                 "well-structured documents for any given topic.",
    name="DocumentPublisher"
)

# One line to make the agent durable with serverless hosting
app = AgentFunctionApp(agents=[agent])

var endpoint = Environment.GetEnvironmentVariable("AZURE_OPENAI_ENDPOINT")
    ?? throw new InvalidOperationException("AZURE_OPENAI_ENDPOINT is not set.");
var deploymentName = Environment.GetEnvironmentVariable("AZURE_OPENAI_DEPLOYMENT_NAME")
    ?? throw new InvalidOperationException("AZURE_OPENAI_DEPLOYMENT_NAME is not set.");

AIAgent agent = new AzureOpenAIClient(new Uri(endpoint), new DefaultAzureCredential())
    .GetChatClient(deploymentName)
    .AsAIAgent(
        instructions: "You are a professional content writer who creates engaging, "
                    + "well-structured documents for any given topic.",
        name: "DocumentPublisher");

// Host the agent with Durable Task Scheduler
string connectionString = "Endpoint=http://localhost:8080;TaskHub=default;Authentication=None";

IHost host = Host.CreateDefaultBuilder(args)
    .ConfigureServices(services =>
    {
        services.ConfigureDurableAgents(
            options => options.AddAIAgent(agent),
            workerBuilder: builder => builder.UseDurableTaskScheduler(connectionString),
            clientBuilder: builder => builder.UseDurableTaskScheduler(connectionString));
    })
    .Build();

await host.StartAsync();

from agent_framework.azure import AzureOpenAIChatClient, DurableAIAgentWorker
from azure.identity import AzureCliCredential
from durabletask.azuremanaged.worker import DurableTaskSchedulerWorker

agent = AzureOpenAIChatClient(credential=AzureCliCredential()).as_agent(
    name="DocumentPublisher",
    instructions="You are a professional content writer who creates engaging, "
                 "well-structured documents for any given topic.",
)

# Create a worker connected to the Durable Task Scheduler
worker = DurableTaskSchedulerWorker(
    host_address="http://localhost:8080",
    secure_channel=False,
    taskhub="default",
)

# Register the agent and start processing
agent_worker = DurableAIAgentWorker(worker)
agent_worker.add_agent(agent)
worker.start()

[Function(nameof(DocumentPublishingOrchestration))]
public async Task<string> DocumentPublishingOrchestration(
    [OrchestrationTrigger] TaskOrchestrationContext context)
{
    var docRequest = context.GetInput<DocumentRequest>();

    DurableAIAgent researchAgent = context.GetAgent("ResearchAgent");
    DurableAIAgent writerAgent = context.GetAgent("DocumentPublisherAgent");

    // Step 1: Research the topic
    AgentResponse<ResearchResult> researchResult = await researchAgent
        .RunAsync<ResearchResult>(
            $"Research the following topic: {docRequest.Topic}");

    // Step 2: Write the document using the research findings
    AgentResponse<DocumentResponse> document = await writerAgent
        .RunAsync<DocumentResponse>(
            $"""Create a document about {docRequest.Topic}.
            Research findings: {researchResult.Result.Findings}""");

    // Step 3: Publish
    return await context.CallActivityAsync<string>(
        nameof(PublishDocument),
        new { docRequest.Topic, document.Result.Text });
}

@app.orchestration_trigger(context_name="context")
def document_publishing_orchestration(context: DurableOrchestrationContext):
    doc_request = context.get_input()

    research_agent = app.get_agent(context, "ResearchAgent")
    writer_agent = app.get_agent(context, "DocumentPublisherAgent")

    research_session = research_agent.create_session()
    writer_session = writer_agent.create_session()

    # Step 1: Research the topic
    research_result = yield research_agent.run(
        messages=f"Research the following topic: {doc_request['topic']}",
        session=research_session,
    )

    # Step 2: Write the document using the research findings
    document = yield writer_agent.run(
        messages=f"""Create a document about {doc_request['topic']}.
        Research findings: {research_result.text}""",
        session=writer_session,
    )

    # Step 3: Publish
    return (yield context.call_activity("publish_document", {
        "title": doc_request["topic"],
        "content": document.text
    }))

static async Task<string> DocumentPublishingOrchestration(
    TaskOrchestrationContext context, DocumentRequest docRequest)
{
    DurableAIAgent researchAgent = context.GetAgent("ResearchAgent");
    DurableAIAgent writerAgent = context.GetAgent("DocumentPublisherAgent");

    // Step 1: Research the topic
    AgentResponse<ResearchResult> researchResult = await researchAgent
        .RunAsync<ResearchResult>(
            $"Research the following topic: {docRequest.Topic}");

    // Step 2: Write the document using the research findings
    AgentResponse<DocumentResponse> document = await writerAgent
        .RunAsync<DocumentResponse>(
            $"""Create a document about {docRequest.Topic}.
            Research findings: {researchResult.Result.Findings}""");

    // Step 3: Publish
    return await context.CallActivityAsync<string>(
        nameof(PublishDocument),
        new { docRequest.Topic, document.Result.Text });
}

from agent_framework.azure import DurableAIAgentOrchestrationContext

def document_publishing_orchestration(ctx, doc_request: dict):
    agent_context = DurableAIAgentOrchestrationContext(ctx)

    research = agent_context.get_agent("ResearchAgent")
    writer = agent_context.get_agent("DocumentPublisherAgent")

    research_session = research.create_session()
    writer_session = writer.create_session()

    # Step 1: Research the topic
    research_result = yield research.run(
        messages=f"Research the following topic: {doc_request['topic']}",
        session=research_session,
    )

    # Step 2: Write the document using the research findings
    document = yield writer.run(
        messages=f"""Create a document about {doc_request['topic']}.
        Research findings: {research_result.text}""",
        session=writer_session,
    )

    # Step 3: Publish
    return (yield ctx.call_activity(publish_document, input={
        "title": doc_request["topic"],
        "content": document.text
    }))

OrderLookup orderLookup = new();
OrderCancel orderCancel = new();
SendEmail sendEmail = new();

Workflow cancelOrder = new WorkflowBuilder(orderLookup)
    .WithName("CancelOrder")
    .WithDescription("Cancel an order and notify the customer")
    .AddEdge(orderLookup, orderCancel)
    .AddEdge(orderCancel, sendEmail)
    .Build();

using IHost app = FunctionsApplication
    .CreateBuilder(args)
    .ConfigureFunctionsWebApplication()
    .ConfigureDurableWorkflows(workflows => workflows.AddWorkflows(cancelOrder))
    .Build();
app.Run();

Die OrderLookup, OrderCancel und SendEmail-Executoren sind Standardmäßige Microsoft Agent Framework-Executoren ohne durable-spezifischen Code. Vollständige Implementierungen finden Sie unter samples auf GitHub.

from agent_framework import Workflow, WorkflowBuilder, WorkflowContext, executor
from agent_framework.azure import AgentFunctionApp

@executor(id="store_email")
async def store_email(email_text: str, ctx: WorkflowContext) -> None:
    ctx.set_state("current_email", email_text)
    await ctx.send_message(email_text)

@executor(id="process_email")
async def process_email(email_text: str, ctx: WorkflowContext) -> None:
    result = f"Processed: {email_text[:50]}..."
    await ctx.send_message(result)

@executor(id="finalize")
async def finalize(result: str, ctx: WorkflowContext[None, str]) -> None:
    await ctx.yield_output(f"Complete: {result}")

workflow = (
    WorkflowBuilder(start_executor=store_email)
    .add_edge(store_email, process_email)
    .add_edge(process_email, finalize)
    .build()
)

app = AgentFunctionApp(workflow=workflow)

ChatClient chatClient = new AzureOpenAIClient(
    new Uri(endpoint), new DefaultAzureCredential()).GetChatClient(deploymentName);

AIAgent physicist = chatClient.AsAIAgent(
    "You are a physics expert. Be concise (2-3 sentences).", "Physicist");
AIAgent chemist = chatClient.AsAIAgent(
    "You are a chemistry expert. Be concise (2-3 sentences).", "Chemist");

ParseQuestionExecutor parseQuestion = new();
AggregatorExecutor aggregator = new();

Workflow workflow = new WorkflowBuilder(parseQuestion)
    .WithName("ExpertReview")
    .AddFanOutEdge(parseQuestion, [physicist, chemist])
    .AddFanInBarrierEdge([physicist, chemist], aggregator)
    .Build();

using IHost app = FunctionsApplication
    .CreateBuilder(args)
    .ConfigureFunctionsWebApplication()
    .ConfigureDurableWorkflows(workflows => workflows.AddWorkflows(workflow))
    .Build();
app.Run();

Die ParseQuestionExecutor und AggregatorExecutor sind Standard-Microsoft Agent Framework-Executoren ohne dauerhaften Code. Vollständige Implementierungen finden Sie unter samples auf GitHub.

from agent_framework import (
    Agent, AgentExecutorResponse, Workflow,
    WorkflowBuilder, WorkflowContext, executor,
)
from agent_framework.azure import AgentFunctionApp

chat_client = ...  # OpenAIChatCompletionClient or FoundryChatClient

sentiment_agent = Agent(
    client=chat_client,
    name="SentimentAnalysisAgent",
    instructions="You are a sentiment analysis expert. Analyze the sentiment of the given text.",
)

keyword_agent = Agent(
    client=chat_client,
    name="KeywordExtractionAgent",
    instructions="You are a keyword extraction expert. Extract important keywords from the given text.",
)

@executor(id="input_router")
async def input_router(doc: str, ctx: WorkflowContext) -> None:
    await ctx.send_message(doc)

@executor(id="prepare_for_output")
async def prepare_for_output(
    analyses: list[AgentExecutorResponse], ctx: WorkflowContext[None, str]
) -> None:
    parts = [f"[{a.executor_id}]: {a.agent_response.text}" for a in analyses]
    await ctx.yield_output("\n\n".join(parts))

workflow = (
    WorkflowBuilder(start_executor=input_router)
    .add_fan_out_edges(source=input_router, targets=[sentiment_agent, keyword_agent])
    .add_fan_in_edges(sources=[sentiment_agent, keyword_agent], target=prepare_for_output)
    .build()
)

app = AgentFunctionApp(workflow=workflow)

AIAgent spamDetector = chatClient.AsAIAgent(
    "You are a spam detection assistant. Return JSON with is_spam (bool) and reason (string).",
    "SpamDetectionAgent");
AIAgent emailAssistant = chatClient.AsAIAgent(
    "You are an email assistant. Draft a professional response.",
    "EmailAssistantAgent");

SpamHandlerExecutor spamHandler = new();
EmailSenderExecutor emailSender = new();

Workflow workflow = new WorkflowBuilder(spamDetector)
    .WithName("EmailClassification")
    .AddSwitchCaseEdgeGroup(spamDetector, [
        new Case(condition: IsSpamDetected, target: spamHandler),
        new Default(target: emailAssistant),
    ])
    .AddEdge(emailAssistant, emailSender)
    .Build();

using IHost app = FunctionsApplication
    .CreateBuilder(args)
    .ConfigureFunctionsWebApplication()
    .ConfigureDurableWorkflows(workflows => workflows.AddWorkflows(workflow))
    .Build();
app.Run();

Die SpamHandlerExecutor und EmailSenderExecutor sind Standard-Microsoft Agent Framework-Executoren ohne dauerhaften Code. Vollständige Implementierungen finden Sie unter samples auf GitHub.

from agent_framework import (
    Agent, AgentExecutorResponse, Case, Default, Executor,
    Workflow, WorkflowBuilder, WorkflowContext, handler,
)
from agent_framework.azure import AgentFunctionApp
from pydantic import BaseModel

class SpamDetectionResult(BaseModel):
    is_spam: bool
    reason: str

chat_client = ...  # FoundryChatClient or OpenAIChatCompletionClient

spam_agent = Agent(
    client=chat_client,
    name="SpamDetectionAgent",
    instructions="You are a spam detection assistant. Return JSON with is_spam and reason.",
    default_options={"response_format": SpamDetectionResult},
)

email_agent = Agent(
    client=chat_client,
    name="EmailAssistantAgent",
    instructions="You are an email assistant that drafts professional responses.",
)

class SpamHandlerExecutor(Executor):
    @handler
    async def handle(self, response: AgentExecutorResponse, ctx: WorkflowContext[None, str]) -> None:
        text = response.agent_response.text
        result = SpamDetectionResult.model_validate_json(text)
        await ctx.yield_output(f"Email marked as spam: {result.reason}")

class EmailSenderExecutor(Executor):
    @handler
    async def handle(self, response: AgentExecutorResponse, ctx: WorkflowContext[None, str]) -> None:
        await ctx.yield_output(f"Email sent: {response.agent_response.text}")

def is_spam_detected(message) -> bool:
    if not isinstance(message, AgentExecutorResponse):
        return False
    result = SpamDetectionResult.model_validate_json(message.agent_response.text)
    return result.is_spam

spam_handler = SpamHandlerExecutor(id="spam_handler")
email_sender = EmailSenderExecutor(id="email_sender")

workflow = (
    WorkflowBuilder(start_executor=spam_agent)
    .add_switch_case_edge_group(spam_agent, [
        Case(condition=is_spam_detected, target=spam_handler),
        Default(target=email_agent),
    ])
    .add_edge(email_agent, email_sender)
    .build()
)

app = AgentFunctionApp(workflow=workflow)

CreateApprovalRequest createRequest = new();
RequestPort<ApprovalRequest, ApprovalResponse> managerApproval =
    RequestPort.Create<ApprovalRequest, ApprovalResponse>("ManagerApproval");
PrepareFinanceReview prepareFinanceReview = new();
RequestPort<ApprovalRequest, ApprovalResponse> budgetApproval =
    RequestPort.Create<ApprovalRequest, ApprovalResponse>("BudgetApproval");
RequestPort<ApprovalRequest, ApprovalResponse> complianceApproval =
    RequestPort.Create<ApprovalRequest, ApprovalResponse>("ComplianceApproval");
ExpenseReimburse reimburse = new();

Workflow expenseApproval = new WorkflowBuilder(createRequest)
    .WithName("ExpenseReimbursement")
    .WithDescription("Expense reimbursement with manager and parallel finance approvals")
    .AddEdge(createRequest, managerApproval)
    .AddEdge(managerApproval, prepareFinanceReview)
    .AddFanOutEdge(prepareFinanceReview, [budgetApproval, complianceApproval])
    .AddFanInBarrierEdge([budgetApproval, complianceApproval], reimburse)
    .Build();

using IHost app = FunctionsApplication
    .CreateBuilder(args)
    .ConfigureFunctionsWebApplication()
    .ConfigureDurableWorkflows(workflows =>
        workflows.AddWorkflow(expenseApproval, exposeStatusEndpoint: true))
    .Build();
app.Run();

Das Framework generiert automatisch drei HTTP-Endpunkte für HITL-Interaktion.

• POST /api/workflows/{name}/run : Starten des Workflows
• GET /api/workflows/{name}/status/{id} : Status überprüfen und ausstehende Freigaben
• POST /api/workflows/{name}/respond/{id} : Genehmigungsantwort senden, um fortzunehmen

Die folgenden Datensatztypen definieren die Daten, die über den Workflow fließen:

public record ApprovalRequest(string ExpenseId, decimal Amount, string EmployeeName);
public record ApprovalResponse(bool Approved, string? Comments);

Die CreateApprovalRequest, PrepareFinanceReview und ExpenseReimburse-Executoren sind Standardmäßige Microsoft Agent Framework-Executoren ohne durable-spezifischen Code. Vollständige Implementierungen finden Sie unter samples auf GitHub.

from agent_framework import (
    Agent, Executor, Workflow, WorkflowBuilder,
    WorkflowContext, handler, response_handler,
)
from agent_framework.azure import AgentFunctionApp
from pydantic import BaseModel

class HumanApprovalResponse(BaseModel):
    approved: bool
    reviewer_notes: str = ""

chat_client = ...  # FoundryChatClient or OpenAIChatCompletionClient

content_analyzer_agent = Agent(
    client=chat_client,
    name="ContentAnalyzerAgent",
    instructions="You are a content moderation assistant. Analyze content for policy compliance.",
)

class InputRouterExecutor(Executor):
    def __init__(self):
        super().__init__(id="input_router")

    @handler
    async def route_input(self, input_text: str, ctx: WorkflowContext) -> None:
        await ctx.send_message(input_text)

class HumanReviewExecutor(Executor):
    def __init__(self):
        super().__init__(id="human_review_executor")

    @handler
    async def request_review(self, data, ctx: WorkflowContext) -> None:
        approval_request = {
            "content": data,
            "prompt": "Please approve or reject this content.",
        }
        await ctx.request_info(
            request_data=approval_request,
            response_type=HumanApprovalResponse,
        )

    @response_handler
    async def handle_approval_response(self, original_request, response, ctx: WorkflowContext) -> None:
        status = "approved" if response.approved else "rejected"
        await ctx.send_message(f"Content {status}: {response.reviewer_notes}")

class PublishExecutor(Executor):
    def __init__(self):
        super().__init__(id="publish_executor")

    @handler
    async def handle_result(self, result: str, ctx: WorkflowContext[None, str]) -> None:
        await ctx.yield_output(result)

input_router = InputRouterExecutor()
human_review = HumanReviewExecutor()
publish = PublishExecutor()

workflow = (
    WorkflowBuilder(start_executor=input_router)
    .add_edge(input_router, content_analyzer_agent)
    .add_edge(content_analyzer_agent, human_review)
    .add_edge(human_review, publish)
    .build()
)

app = AgentFunctionApp(workflow=workflow)