[AG-UI详解-06]AG-UI针对MAF的服务端实现

JaydenAI2026-07-25 14:15

在AG-UI详解-04：如何构建一个兼容AG-UI协议的Agent中，我通过在ASP.NET Core应用中注册了一个绑定到IChatClient的路由中间件，构建了一个兼容AG-UI协议的Agent后端。虽然AG-UI协议定义的很多功能未曾实现，但是它与MAF与AG-UI集成的实现原理本质上是一致的。现在我们来看看微软官方是如何将一个MAF AIAgent暴露成一个兼容AG-UI的后端服务的。

1. 基于AG-UI的路由注册

和我的做法一样，微软官方也是通过在ASP.NET Core应用中注册相应路由终结点的方式来构建兼容AG-UI兼容Agent后端服务的。相应的API定义在Microsoft.Agents.AI.Hosting.AGUI.AspNetCore这个NuGet包中。值得一提的是，这部分的内容在最近发生了重大的改变，主要体现在针对AG-UI基础功能的依赖全部转移到AG-UI官方SDK上（之前微软自行实现了一套）。包含最新API的NuGet尚未发布，所以如下的实例使用的是直接下载的源代码，而不是应用的NuGet包。

路由注册实现在这个MapAGUIServer扩展方法上，和我在AG-UI详解-04：如何构建一个兼容AG-UI协议的Agent中演示的不同之处在于，这里绑定的是一个AIAgent对象，而我使用的是一个IChatClient对象。另一个AddAGUIServer扩展方法用于注册所需的依赖服务，目前只要体现在针对AG-UI兼容JsonOptions的设置。

csharp 复制代码 
public static IServiceCollection AddAGUIServer(this IServiceCollection services)；
public static IEndpointConventionBuilder MapAGUIServer(
    this IEndpointRouteBuilder endpoints,
    [StringSyntax("route")] string pattern,
    AIAgent aiAgent)；

我们指定的AIAgent最终将会使用如下这个AIHostAgent类型的Agent中间件进行封装，其目的在于利用注册的AgentSessionStore上下Session的持久化。如果没有注册AgentSessionStore，默认会创建一个不具有持久化能力的NoopAgentSessionStore对象，并以IsolationKeyScopedAgentSessionStore作进一步封装。

csharp 复制代码 
public class AIHostAgent : DelegatingAIAgent
{
    private readonly AgentSessionStore _sessionStore;
    public AIHostAgent(AIAgent innerAgent, AgentSessionStore sessionStore)
        : base(innerAgent)
        =>_sessionStore = sessionStore
    public ValueTask<AgentSession> GetOrCreateSessionAsync(string conversationId, CancellationToken cancellationToken = default)
        => _sessionStore.GetSessionAsync(this.InnerAgent, conversationId, cancellationToken);
    public ValueTask SaveSessionAsync(string conversationId, AgentSession session, CancellationToken cancellationToken = default)
        =>_sessionStore.SaveSessionAsync(this.InnerAgent, conversationId, session, cancellationToken);
}

public abstract class AgentSessionStore
{
	public abstract ValueTask SaveSessionAsync(
        AIAgent agent, 
        string conversationId, 
        AgentSession session, 
        CancellationToken cancellationToken = default);

	public abstract ValueTask<AgentSession> GetSessionAsync(
        AIAgent agent, 
        string conversationId, 
        CancellationToken cancellationToken = default);
}

总体来说，注册的路由终结点会采用如下的流程来处理请求：

  • 提取作为输入的RunAgentInput作为当前请求上下文；
  • 将RunAgentInput携带的ThreadId（如果没有就创建一个新的）作为输入调用AIHostAgent的GetOrCreateSessionAsync方法创建或者提取AgentSession对象；
  • 将RunAgentInput携带的AGUIMessage列表转换成ChatMessage列表，连同AgentSession作为参数调用AIHostAgent的RunStreamingAsync方法。
  • 将作为返回值的IAsyncEnumerable<AgentResponseUpdate>转换成IAsyncEnumerable<ChatResponseUpdate>；
  • 再将IAsyncEnumerable<ChatResponseUpdate>转换成IAsyncEnumerable<BaseEvent>
  • 按照SSE协议将BaseEvent序列化成JSON依次输出。

上述流程作核心的部分莫过于将IAsyncEnumerable<ChatResponseUpdate>转换成IAsyncEnumerable<BaseEvent>，AG-UI详解-04：如何构建一个兼容AG-UI协议的Agent也提供了对这一流程的模拟，只是我作为演示仅仅考虑了少数几种AIContent类型而已。接下来我们通过实例演示的方式，看看针对不同应用场景生成的AIContent最终都转换成怎样的AG-UI事件。

2. 常规调用

我们看看最常规的调用会生成怎样的AG-UI事件。如下面的演示程序所示，我们针对OpenAIClient创建了一个AIAgent，并将其作为参数调用了上述的MapAGUIServer扩展方法注册了针对根路径的AG-UI路由。我们直接创建对应的RunAgentInput对象作为输入，并利用HttpClient将输入序列化成JSON发送到AG-UI服务应用的监听地址，并将相应的内容直接输出到控制台上。

csharp 复制代码 
using AGUI.Abstractions;
using DotNetEnv;
using Microsoft.Agents.AI.Hosting.AGUI.AspNetCore;
using Microsoft.Extensions.AI;
using OpenAI;
using OpenAI.Chat;
using System.ClientModel;

Env.Load();
var endpoint = Environment.GetEnvironmentVariable("OPENAI_BASE_URL")!;
var apiKey = Environment.GetEnvironmentVariable("OPENAI_API_KEY")!;
var model = Environment.GetEnvironmentVariable("MODEL")!;

var builder = WebApplication.CreateBuilder(args);
builder.WebHost.UseUrls("http://localhost:5566");
builder.Services.AddHttpClient().AddLogging();
builder.Services.AddAGUIServer();
var app = builder.Build();

var agent = new OpenAIClient(new ApiKeyCredential(apiKey), new OpenAIClientOptions { Endpoint = new Uri(endpoint) })
    .GetChatClient(model)
    .AsIChatClient()
    .AsAIAgent();
app.MapAGUIServer("/", agent);
await app.StartAsync();

var httpClient = new HttpClient();
var input = new RunAgentInput
{
    ThreadId = "thread-123",
    RunId = $"run-123",
    Messages = [new AGUIUserMessage { Content = "三苏故里是哪个城市？" }]
};

var response = await httpClient.PostAsJsonAsync("http://localhost:5566", input);
var content = await response.Content.ReadAsStringAsync();
Console.WriteLine(content);
Console.ReadLine();

输出

复制代码 
data: {"type":"RUN_STARTED","threadId":"thread-123","runId":"run-123"}

data: {"type":"TEXT_MESSAGE_START","messageId":"chatcmpl-E4mkT04E7giWgNq5XOX6lHmALqYuo","role":"assistant","name":"","rawEvent":{"authorName":"","role":"assistant","contents":[{"$type":"text","text":"","annotations":null,"additionalProperties":null}],"additionalProperties":null,"responseId":"chatcmpl-E4mkT04E7giWgNq5XOX6lHmALqYuo","messageId":"chatcmpl-E4mkT04E7giWgNq5XOX6lHmALqYuo","conversationId":null,"createdAt":"2026-07-23T12:23:29+00:00","finishReason":null,"modelId":"","continuationToken":null}}

data: {"type":"TEXT_MESSAGE_CONTENT","messageId":"chatcmpl-E4mkT04E7giWgNq5XOX6lHmALqYuo","delta":"四川","rawEvent":{"authorName":"","role":"assistant","contents":[{"$type":"text","text":"四川","annotations":null,"additionalProperties":null}],"additionalProperties":null,"responseId":"chatcmpl-E4mkT04E7giWgNq5XOX6lHmALqYuo","messageId":"chatcmpl-E4mkT04E7giWgNq5XOX6lHmALqYuo","conversationId":null,"createdAt":"2026-07-23T12:23:29+00:00","finishReason":null,"modelId":"","continuationToken":null}}

data: {"type":"TEXT_MESSAGE_CONTENT","messageId":"chatcmpl-E4mkT04E7giWgNq5XOX6lHmALqYuo","delta":"眉","rawEvent":{"authorName":"","role":"assistant","contents":[{"$type":"text","text":"眉","annotations":null,"additionalProperties":null}],"additionalProperties":null,"responseId":"chatcmpl-E4mkT04E7giWgNq5XOX6lHmALqYuo","messageId":"chatcmpl-E4mkT04E7giWgNq5XOX6lHmALqYuo","conversationId":null,"createdAt":"2026-07-23T12:23:29+00:00","finishReason":null,"modelId":"","continuationToken":null}}

data: {"type":"TEXT_MESSAGE_CONTENT","messageId":"chatcmpl-E4mkT04E7giWgNq5XOX6lHmALqYuo","delta":"山","rawEvent":{"authorName":"","role":"assistant","contents":[{"$type":"text","text":"山","annotations":null,"additionalProperties":null}],"additionalProperties":null,"responseId":"chatcmpl-E4mkT04E7giWgNq5XOX6lHmALqYuo","messageId":"chatcmpl-E4mkT04E7giWgNq5XOX6lHmALqYuo","conversationId":null,"createdAt":"2026-07-23T12:23:29+00:00","finishReason":null,"modelId":"","continuationToken":null}}

data: {"type":"TEXT_MESSAGE_CONTENT","messageId":"chatcmpl-E4mkT04E7giWgNq5XOX6lHmALqYuo","delta":"。","rawEvent":{"authorName":"","role":"assistant","contents":[{"$type":"text","text":"。","annotations":null,"additionalProperties":null}],"additionalProperties":null,"responseId":"chatcmpl-E4mkT04E7giWgNq5XOX6lHmALqYuo","messageId":"chatcmpl-E4mkT04E7giWgNq5XOX6lHmALqYuo","conversationId":null,"createdAt":"2026-07-23T12:23:29+00:00","finishReason":null,"modelId":"","continuationToken":null}}

data: {"type":"TEXT_MESSAGE_END","messageId":"chatcmpl-E4mkT04E7giWgNq5XOX6lHmALqYuo"}

data: {"type":"RUN_FINISHED","threadId":"thread-123","runId":"run-123","outcome":{"type":"success"}}

整个输出体现了如下的AG-UI事件流：

  • RUN_STARTED： Agent开始执行；
  • TEXT_MESSAGE_START ：一条ID为chatcmpl-E4mkT04E7giWgNq5XOX6lHmALqYuo开始传输；
  • TEXT_MESSAGE_CONTENT：消息的内容以增量的形式实时传输；
  • TEXT_MESSAGE_END：本条消息传输完毕；
  • RUN_FINISHED： Agent执行正常完成。

3. 工具调用（Agent端）

我们进一步来看看如果涉及到工具调用，又会输出怎样的AG-UI事件。如下面的演示程序所示，我们创建了一个用来返回指定城市天气信息的工具函数GetWeather，并注册到创建的AIAgent对象上。并且提了一个必须调用此工具才能回答的问题:目前眉山下雨吗？你只需要回答是或者否。

csharp 复制代码 
using AGUI.Abstractions;
using DotNetEnv;
using Microsoft.Agents.AI.Hosting.AGUI.AspNetCore;
using Microsoft.Extensions.AI;
using OpenAI;
using System.ClientModel;
using System.ComponentModel;

Env.Load();
var endpoint = Environment.GetEnvironmentVariable("OPENAI_BASE_URL")!;
var apiKey = Environment.GetEnvironmentVariable("OPENAI_API_KEY")!;
var model = Environment.GetEnvironmentVariable("MODEL")!;

var builder = WebApplication.CreateBuilder(args);
builder.WebHost.UseUrls("http://localhost:5566");
builder.Services.AddHttpClient().AddLogging();
builder.Services.AddAGUIServer();
var app = builder.Build();

var tool = AIFunctionFactory.Create(GetWeather, nameof(GetWeather));
var agent = new OpenAIClient(new ApiKeyCredential(apiKey), new OpenAIClientOptions { Endpoint = new Uri(endpoint) })
    .GetChatClient(model)
    .AsIChatClient()
    .AsAIAgent(tools: [tool]);
app.MapAGUIServer("/", agent);
await app.StartAsync();

var httpClient = new HttpClient();
var input = new RunAgentInput
{
    ThreadId = "thread-123",
    RunId = $"run-123",
    Messages = [new AGUIUserMessage { Content = "目前眉山下雨吗？你只需要回答是或者否。" }]
};

var response = await httpClient.PostAsJsonAsync("http://localhost:5566", input);
var content = await response.Content.ReadAsStringAsync();
Console.WriteLine(content);
Console.ReadLine();

[Description("获取指定城市天气")]
static string GetWeather([Description("城市名称")] string city) => $"晴，气温55摄氏度";

输出：

复制代码 
data: {"type":"RUN_STARTED","threadId":"thread-123","runId":"run-123"}

data: {"type":"TOOL_CALL_START","parentMessageId":"","toolCallId":"call_czyxLE8NKL2q4h116WmiTYxB","toolCallName":"GetWeather","rawEvent":{"authorName":"","role":"assistant","contents":[{"$type":"functionCall","name":"GetWeather","arguments":{"city":"眉山"},"informationalOnly":false,"callId":"call_czyxLE8NKL2q4h116WmiTYxB","annotations":null,"additionalProperties":null}],"additionalProperties":null,"responseId":"","messageId":"","conversationId":null,"createdAt":"1970-01-01T00:00:00+00:00","finishReason":"tool_calls","modelId":"","continuationToken":null}}

data: {"type":"TOOL_CALL_ARGS","toolCallId":"call_czyxLE8NKL2q4h116WmiTYxB","delta":"{\"city\":\"眉山\"}","rawEvent":{"authorName":"","role":"assistant","contents":[{"$type":"functionCall","name":"GetWeather","arguments":{"city":"眉山"},"informationalOnly":false,"callId":"call_czyxLE8NKL2q4h116WmiTYxB","annotations":null,"additionalProperties":null}],"additionalProperties":null,"responseId":"","messageId":"","conversationId":null,"createdAt":"1970-01-01T00:00:00+00:00","finishReason":"tool_calls","modelId":"","continuationToken":null}}

data: {"type":"TOOL_CALL_END","toolCallId":"call_czyxLE8NKL2q4h116WmiTYxB","rawEvent":{"authorName":"","role":"assistant","contents":[{"$type":"functionCall","name":"GetWeather","arguments":{"city":"眉山"},"informationalOnly":false,"callId":"call_czyxLE8NKL2q4h116WmiTYxB","annotations":null,"additionalProperties":null}],"additionalProperties":null,"responseId":"","messageId":"","conversationId":null,"createdAt":"1970-01-01T00:00:00+00:00","finishReason":"tool_calls","modelId":"","continuationToken":null}}

data: {"type":"TOOL_CALL_RESULT","messageId":"call_czyxLE8NKL2q4h116WmiTYxB","toolCallId":"call_czyxLE8NKL2q4h116WmiTYxB","content":"\"晴，气温55摄氏度\"","rawEvent":{"authorName":"","role":"tool","contents":[{"$type":"functionResult","result":"晴，气温55摄氏度","callId":"call_czyxLE8NKL2q4h116WmiTYxB","annotations":null,"additionalProperties":null}],"additionalProperties":null,"responseId":"9011115d625947f1b24a3a536b8d424a","messageId":"9011115d625947f1b24a3a536b8d424a","conversationId":null,"createdAt":"2026-07-23T13:18:20.3295097+00:00","finishReason":null,"modelId":null,"continuationToken":null}}

data: {"type":"TEXT_MESSAGE_START","messageId":"chatcmpl-E4nbYTrEOsAv8gFlv2xEaPbEPzygG","role":"assistant","name":"","rawEvent":{"authorName":"","role":"assistant","contents":[{"$type":"text","text":"","annotations":null,"additionalProperties":null}],"additionalProperties":null,"responseId":"chatcmpl-E4nbYTrEOsAv8gFlv2xEaPbEPzygG","messageId":"chatcmpl-E4nbYTrEOsAv8gFlv2xEaPbEPzygG","conversationId":null,"createdAt":"2026-07-23T13:18:20+00:00","finishReason":null,"modelId":"","continuationToken":null}}

data: {"type":"TEXT_MESSAGE_CONTENT","messageId":"chatcmpl-E4nbYTrEOsAv8gFlv2xEaPbEPzygG","delta":"否","rawEvent":{"authorName":"","role":"assistant","contents":[{"$type":"text","text":"否","annotations":null,"additionalProperties":null}],"additionalProperties":null,"responseId":"chatcmpl-E4nbYTrEOsAv8gFlv2xEaPbEPzygG","messageId":"chatcmpl-E4nbYTrEOsAv8gFlv2xEaPbEPzygG","conversationId":null,"createdAt":"2026-07-23T13:18:20+00:00","finishReason":null,"modelId":"","continuationToken":null}}

data: {"type":"TEXT_MESSAGE_END","messageId":"chatcmpl-E4nbYTrEOsAv8gFlv2xEaPbEPzygG"}

data: {"type":"RUN_FINISHED","threadId":"thread-123","runId":"run-123","outcome":{"type":"success"}}

本地调用体现了如下的AG-UI事件流：

  • RUN_STARTED：Agent开始执行；
  • TOOL_CALL_START ：开始传输针对GetWeather的工具调用；
  • TOOL_CALL_ARGS ：传入工具调用的参数（"city":"眉山"）；
  • TOOL_CALL_END：工具调用传输结束；
  • TOOL_CALL_RESULT：传输工具执行的结果（"晴，气温55摄氏度"）；
  • TEXT_MESSAGE_START:开始传输LLM最终的答复的文本消息；
  • TEXT_MESSAGE_CONTENT：传入消息的内容（"否"）
  • TEXT_MESSAGE_END:文本消息传输结束；
  • RUN_FINISHED：Agent执行正常完成。

4. 推理

我们在上面这个演示实例中进一步引入推理。由于我们无法控制LLM输出推理的内容，所以我们定义了如下这个名为ReasoningChatClient的ChatClient中间件，使用它来输出一个表示推理的TextReasoningContent对象。我们将它注册在AIAgent的ChatClient管道中。

csharp 复制代码 
var agent = new OpenAIClient(new ApiKeyCredential(apiKey), new OpenAIClientOptions { Endpoint = new Uri(endpoint) })
    .GetChatClient(model)
    .AsIChatClient()
    .AsBuilder()
    .Use(innerClient=>new ReasoningChatClient(innerClient))
    .Build()
    .AsAIAgent(tools: [tool]);

class ReasoningChatClient(IChatClient innerClient) : DelegatingChatClient(innerClient)
{
    public override async IAsyncEnumerable<ChatResponseUpdate> GetStreamingResponseAsync(
        IEnumerable<ChatMessage> messages, 
        ChatOptions? options = null,
        [EnumeratorCancellation]CancellationToken cancellationToken = default)
    {
        
        await foreach (var update in base.GetStreamingResponseAsync(messages, options, cancellationToken))
        {
            yield return update;
        }
        if (messages.Any(it => it.Role == ChatRole.Tool))
        {
            yield return new ChatResponseUpdate(ChatRole.Assistant, [new TextReasoningContent("极端天气，真实度存疑")]);
        }
    }
}

输出：

复制代码 
data: {"type":"RUN_STARTED","threadId":"thread-123","runId":"run-123"}

data: {"type":"TOOL_CALL_START","parentMessageId":"","toolCallId":"call_uZoyZo4rJYuK7zWmpB65AF3M","toolCallName":"GetWeather","rawEvent":{"authorName":null,"role":"assistant","contents":[{"$type":"functionCall","name":"GetWeather","arguments":{"city":"眉山"},"informationalOnly":false,"callId":"call_uZoyZo4rJYuK7zWmpB65AF3M","annotations":null,"additionalProperties":null}],"additionalProperties":null,"responseId":"","messageId":"","conversationId":null,"createdAt":"1970-01-01T00:00:00+00:00","finishReason":"tool_calls","modelId":"","continuationToken":null}}

data: {"type":"TOOL_CALL_ARGS","toolCallId":"call_uZoyZo4rJYuK7zWmpB65AF3M","delta":"{\"city\":\"眉山\"}","rawEvent":{"authorName":null,"role":"assistant","contents":[{"$type":"functionCall","name":"GetWeather","arguments":{"city":"眉山"},"informationalOnly":false,"callId":"call_uZoyZo4rJYuK7zWmpB65AF3M","annotations":null,"additionalProperties":null}],"additionalProperties":null,"responseId":"","messageId":"","conversationId":null,"createdAt":"1970-01-01T00:00:00+00:00","finishReason":"tool_calls","modelId":"","continuationToken":null}}

data: {"type":"TOOL_CALL_END","toolCallId":"call_uZoyZo4rJYuK7zWmpB65AF3M","rawEvent":{"authorName":null,"role":"assistant","contents":[{"$type":"functionCall","name":"GetWeather","arguments":{"city":"眉山"},"informationalOnly":false,"callId":"call_uZoyZo4rJYuK7zWmpB65AF3M","annotations":null,"additionalProperties":null}],"additionalProperties":null,"responseId":"","messageId":"","conversationId":null,"createdAt":"1970-01-01T00:00:00+00:00","finishReason":"tool_calls","modelId":"","continuationToken":null}}

data: {"type":"TOOL_CALL_RESULT","messageId":"call_uZoyZo4rJYuK7zWmpB65AF3M","toolCallId":"call_uZoyZo4rJYuK7zWmpB65AF3M","content":"\"晴，气温55摄氏度\"","rawEvent":{"authorName":null,"role":"tool","contents":[{"$type":"functionResult","result":"晴，气温55摄氏度","callId":"call_uZoyZo4rJYuK7zWmpB65AF3M","annotations":null,"additionalProperties":null}],"additionalProperties":null,"responseId":"0abc336f1844454a9d4bac380c590723","messageId":"0abc336f1844454a9d4bac380c590723","conversationId":null,"createdAt":"2026-07-23T13:40:18.3569533+00:00","finishReason":null,"modelId":null,"continuationToken":null}}

data: {"type":"TEXT_MESSAGE_START","messageId":"chatcmpl-E4nwojNkIkTtwWdR2f3p2umffcxgU","role":"assistant","rawEvent":{"authorName":null,"role":"assistant","contents":[{"$type":"text","text":"","annotations":null,"additionalProperties":null}],"additionalProperties":null,"responseId":"chatcmpl-E4nwojNkIkTtwWdR2f3p2umffcxgU","messageId":"chatcmpl-E4nwojNkIkTtwWdR2f3p2umffcxgU","conversationId":null,"createdAt":"2026-07-23T13:40:18+00:00","finishReason":null,"modelId":"","continuationToken":null}}

data: {"type":"TEXT_MESSAGE_CONTENT","messageId":"chatcmpl-E4nwojNkIkTtwWdR2f3p2umffcxgU","delta":"否","rawEvent":{"authorName":null,"role":"assistant","contents":[{"$type":"text","text":"否","annotations":null,"additionalProperties":null}],"additionalProperties":null,"responseId":"chatcmpl-E4nwojNkIkTtwWdR2f3p2umffcxgU","messageId":"chatcmpl-E4nwojNkIkTtwWdR2f3p2umffcxgU","conversationId":null,"createdAt":"2026-07-23T13:40:18+00:00","finishReason":null,"modelId":"","continuationToken":null}}

data: {"type":"TEXT_MESSAGE_END","messageId":"chatcmpl-E4nwojNkIkTtwWdR2f3p2umffcxgU","rawEvent":{"authorName":null,"role":"assistant","contents":[{"$type":"reasoning","text":"极端天气，真实度存疑","protectedData":null,"annotations":null,"additionalProperties":null}],"additionalProperties":null,"responseId":null,"messageId":null,"conversationId":null,"createdAt":null,"finishReason":null,"modelId":null,"continuationToken":null}}

data: {"type":"REASONING_START","messageId":"msg_rTPEmtDr8zDc2X4ZTFH2sk35"}

data: {"type":"REASONING_MESSAGE_START","messageId":"msg_rTPEmtDr8zDc2X4ZTFH2sk35","role":"reasoning"}

data: {"type":"REASONING_MESSAGE_CONTENT","messageId":"msg_rTPEmtDr8zDc2X4ZTFH2sk35","delta":"极端天气，真实度存疑"}

data: {"type":"REASONING_MESSAGE_END","messageId":"msg_rTPEmtDr8zDc2X4ZTFH2sk35"}

data: {"type":"REASONING_END","messageId":"msg_rTPEmtDr8zDc2X4ZTFH2sk35"}

data: {"type":"RUN_FINISHED","threadId":"thread-123","runId":"run-123","outcome":{"type":"success"}}

由于ReasoningChatClient是在LLM的应答之后输出的推理内容，所以会看来基于推理的AG-UI事件出现在事件流的末端：

  • REASONING_START: 开始推理流程；
  • REASONING_MESSAGE_START: 开始传输指定的推理消息；
  • REASONING_MESSAGE_CONTENT: 传输推理消息的内容；
  • REASONING_MESSAGE_END: 指定推理消息传输结束；
  • REASONING_END: 推理流程结束。

我觉得这里的实现可能是有问题的。按照AG-UI协议文档的说明，作为REASONING_START/REASONING_END和REASONING_MESSAGE_START/REASONING_MESSAGE_END分别作为整个推理流程和单条推理消息的边界，所以它们的messageId并不是一回事，我查看了实现代码，貌似他们将两者混为一谈了。

5. 中断

我们最后来看看如果涉及人机交互导致Agent执行中断，AG-UI事件流又会出现什么变化。为此我们使用ApprovalRequiredAIFunction将注册的工具进行封装，这样会因工具审批导致中断。

csharp 复制代码 
var tool = new ApprovalRequiredAIFunction(AIFunctionFactory.Create(GetWeather, nameof(GetWeather)));

输出：

复制代码 
data: {"type":"RUN_STARTED","threadId":"thread-123","runId":"run-123"}

data: {"type":"TOOL_CALL_START","parentMessageId":"","toolCallId":"call_lWBq3b0a2Qs42WXwtnaeyByk","toolCallName":"GetWeather","rawEvent":{"authorName":"","role":"assistant","contents":[{"$type":"toolApprovalRequest","toolCall":{"$type":"functionCall","name":"GetWeather","arguments":{"city":"眉山"},"informationalOnly":false,"callId":"call_lWBq3b0a2Qs42WXwtnaeyByk","annotations":null,"additionalProperties":null},"requestId":"ficc_call_lWBq3b0a2Qs42WXwtnaeyByk","annotations":null,"additionalProperties":null}],"additionalProperties":null,"responseId":"","messageId":"","conversationId":null,"createdAt":"1970-01-01T00:00:00+00:00","finishReason":"tool_calls","modelId":"","continuationToken":null}}

data: {"type":"TOOL_CALL_ARGS","toolCallId":"call_lWBq3b0a2Qs42WXwtnaeyByk","delta":"{\"city\":\"眉山\"}","rawEvent":{"authorName":"","role":"assistant","contents":[{"$type":"toolApprovalRequest","toolCall":{"$type":"functionCall","name":"GetWeather","arguments":{"city":"眉山"},"informationalOnly":false,"callId":"call_lWBq3b0a2Qs42WXwtnaeyByk","annotations":null,"additionalProperties":null},"requestId":"ficc_call_lWBq3b0a2Qs42WXwtnaeyByk","annotations":null,"additionalProperties":null}],"additionalProperties":null,"responseId":"","messageId":"","conversationId":null,"createdAt":"1970-01-01T00:00:00+00:00","finishReason":"tool_calls","modelId":"","continuationToken":null}}

data: {"type":"TOOL_CALL_END","toolCallId":"call_lWBq3b0a2Qs42WXwtnaeyByk","rawEvent":{"authorName":"","role":"assistant","contents":[{"$type":"toolApprovalRequest","toolCall":{"$type":"functionCall","name":"GetWeather","arguments":{"city":"眉山"},"informationalOnly":false,"callId":"call_lWBq3b0a2Qs42WXwtnaeyByk","annotations":null,"additionalProperties":null},"requestId":"ficc_call_lWBq3b0a2Qs42WXwtnaeyByk","annotations":null,"additionalProperties":null}],"additionalProperties":null,"responseId":"","messageId":"","conversationId":null,"createdAt":"1970-01-01T00:00:00+00:00","finishReason":"tool_calls","modelId":"","continuationToken":null}}

data: {"type":"RUN_FINISHED","threadId":"thread-123","runId":"run-123","outcome":{"type":"interrupt","interrupts":[{"id":"ficc_call_lWBq3b0a2Qs42WXwtnaeyByk","reason":"tool_call","message":"Approval required for tool call: GetWeather","toolCallId":"call_lWBq3b0a2Qs42WXwtnaeyByk","responseSchema":{"type":"object","properties":{"approved":{"type":"boolean"}},"required":["approved"]}}]}}

按照AG-UI协议的描述，中断体现在RUN_FINISHED事件的outcome字段上。从最后输出的RUN_FINISHED事件内容可以看出，outcome字段具有如下的设置：

  • type：interrupt
  • interrupts :
    • id: ficc_call_lWBq3b0a2Qs42WXwtnaeyByk
    • reason:tool_call
    • message: Approval required for tool call: GetWeather
    • toolCallId: call_lWBq3b0a2Qs42WXwtnaeyByk
    • responseSchema: {"type":"object","properties":{"approved":{"type":"boolean"}},"required":"approved"}

6. 工具调用（客户端）

我们最后来看看如果工具不是直接注册到承载AIAgent上，都是直接由客户端提供的情况下，按照我们的认知，此时工具调用应该传输给前端应用，并在应用端本地执行。我们看看最终会得到怎样的AG-UI事件流。

为此我们对演示程序做了如下的改动。我们创建的AIAgent不再有任何的工具注册。取而代之的是，我们根据AIFunction创建的AGUITool对象被添加到RunAgentInput对象的Tools列表中。

csharp 复制代码 
using AGUI.Abstractions;
using DotNetEnv;
using Microsoft.Agents.AI.Hosting.AGUI.AspNetCore;
using Microsoft.Extensions.AI;
using OpenAI;
using System.ClientModel;
using System.ComponentModel;
using System.Runtime.CompilerServices;

Env.Load();
var endpoint = Environment.GetEnvironmentVariable("OPENAI_BASE_URL")!;
var apiKey = Environment.GetEnvironmentVariable("OPENAI_API_KEY")!;
var model = Environment.GetEnvironmentVariable("MODEL")!;

var builder = WebApplication.CreateBuilder(args);
builder.WebHost.UseUrls("http://localhost:5566");
builder.Services.AddHttpClient().AddLogging();
builder.Services.AddAGUIServer();
var app = builder.Build();

var agent = new OpenAIClient(new ApiKeyCredential(apiKey), new OpenAIClientOptions { Endpoint = new Uri(endpoint) })
    .GetChatClient(model)
    .AsIChatClient()
    .AsAIAgent();
app.MapAGUIServer("/", agent);
await app.StartAsync();

var httpClient = new HttpClient();
var tool = AIFunctionFactory.Create(GetWeather, nameof(GetWeather));
var aguiTool = new AGUITool
{
    Name = tool.Name,
    Description = tool.Description,
    Parameters = tool.JsonSchema
};
var input = new RunAgentInput
{
    ThreadId = "thread-123",
    RunId = $"run-123",
    Tools = [aguiTool],
    Messages = [new AGUIUserMessage { Content = "目前眉山下雨吗？你只需要回答是或者否。" }]
};

var response = await httpClient.PostAsJsonAsync("http://localhost:5566", input);
var content = await response.Content.ReadAsStringAsync();
Console.WriteLine(content);
Console.ReadLine();

[Description("获取指定城市天气")]
static string GetWeather([Description("城市名称")] string city) => $"晴，气温55摄氏度";

输出：

复制代码 
data: {"type":"RUN_STARTED","threadId":"thread-123","runId":"run-123"}

data: {"type":"TOOL_CALL_START","parentMessageId":"","toolCallId":"call_xG9kRZiGtJpM2LJGHdKEqFUX","toolCallName":"GetWeather","rawEvent":{"authorName":null,"role":"assistant","contents":[{"$type":"functionCall","name":"GetWeather","arguments":{"city":"眉山"},"informationalOnly":false,"callId":"call_xG9kRZiGtJpM2LJGHdKEqFUX","annotations":null,"additionalProperties":null}],"additionalProperties":null,"responseId":"","messageId":"","conversationId":null,"createdAt":"1970-01-01T00:00:00+00:00","finishReason":"tool_calls","modelId":"","continuationToken":null}}

data: {"type":"TOOL_CALL_ARGS","toolCallId":"call_xG9kRZiGtJpM2LJGHdKEqFUX","delta":"{\"city\":\"眉山\"}","rawEvent":{"authorName":null,"role":"assistant","contents":[{"$type":"functionCall","name":"GetWeather","arguments":{"city":"眉山"},"informationalOnly":false,"callId":"call_xG9kRZiGtJpM2LJGHdKEqFUX","annotations":null,"additionalProperties":null}],"additionalProperties":null,"responseId":"","messageId":"","conversationId":null,"createdAt":"1970-01-01T00:00:00+00:00","finishReason":"tool_calls","modelId":"","continuationToken":null}}

data: {"type":"TOOL_CALL_END","toolCallId":"call_xG9kRZiGtJpM2LJGHdKEqFUX","rawEvent":{"authorName":null,"role":"assistant","contents":[{"$type":"functionCall","name":"GetWeather","arguments":{"city":"眉山"},"informationalOnly":false,"callId":"call_xG9kRZiGtJpM2LJGHdKEqFUX","annotations":null,"additionalProperties":null}],"additionalProperties":null,"responseId":"","messageId":"","conversationId":null,"createdAt":"1970-01-01T00:00:00+00:00","finishReason":"tool_calls","modelId":"","continuationToken":null}}

data: {"type":"RUN_FINISHED","threadId":"thread-123","runId":"run-123","outcome":{"type":"success"}}

输出体现的AG-UI事件流：

  • RUN_STARTED： Agent开始执行；
  • TOOL_CALL_START ： 针对客户端工具GetWeather的工具调用开始传输；
  • TOOL_CALL_ARGS ：传输工具调用的参数（{"city":"眉山"}）;
  • RUN_FINISHED：Agent执行正常完成。

此后客户端应该执行工具函数，然后将结果转换成AGUIToolMessage添加在作为输入的RunAgentInput的Messages列表中，再次发起调用。

相关推荐
一包惆怅的辣条2 小时前
谈一下怎么写一个一套可复用的用例改写skill架构
自动化测试·人工智能·ai·skill
JouYY3 小时前
大模型底层学习（二）- 预训练（Pre-training）流程概览
架构·llm·agent
weixin_666593993 小时前
自然资源时空智能体建设方案：从数字化到智能化的演进路径
大数据·人工智能·大模型·云计算·agent·云平台·空间数据库
wWYy.4 小时前
什么是RAG？
人工智能·agent
俊哥V4 小时前
每日 AI 研究简报 · 2026-07-24
人工智能·ai
深念Y4 小时前
CC Switch 显示错误模型名的排查与解决
网关·开源·agent·项目·代理
大公产经晚间消息4 小时前
“胜算未来·2026戈峻思享会”即将登陆大连
大数据·人工智能·ai
武子康5 小时前
同一套 ComfyUI 工作流第二次为什么快一倍：4 类消息误解 + 互斥阶段账本 + 10 单元实验矩阵
人工智能·aigc·agent
热门推荐
01GitHub 镜像站点02如何新建文件夹？ 电脑新建文件夹的4种方法032026年7月AI圈大地震：GPT-5.6被政府限制、Claude入驻Slack、Anthropic自研芯片04AI科技热点日报 | 2026年07月01日05国内可直接用、免费额度/永久免费的大模型API清单（含 SiliconFlow、火山、阿里、智谱、百度、Kimi、DeepSeek、DMXAPI 等）06AI 编程 IDE 全景解析 2026：Agent 全面接管开发链路07幻兽帕鲁 - 服务器管理员权限与 GM 命令完全指南08微信历史版本含下载地址（ Windows PC | 安卓 | MAC ）及设置微信不更新092026 国产 AI 大模型横评：DeepSeek、通义千问、Kimi、文心一言、星火、豆包谁更能打？10【解构】DeepSeek V4 发布：技术报告深度解读 + 横向对比六大开源模型，我们的判断是……