工作流代理

工作流代理以可预测的模式协调多个代理——顺序流水线、并行执行或迭代循环。不同于使用 AI 推理的 LlmAgent，工作流代理遵循确定性的执行路径。

快速开始

创建一个新项目：

cargo new workflow_demo
cd workflow_demo

将依赖添加到 Cargo.toml：

[dependencies]
adk-rust = "2.0.0"
tokio = { version = "1.40", features = ["full"] }
dotenvy = "0.15"

创建 .env：

echo 'GOOGLE_API_KEY=your-api-key' > .env

---

SequentialAgent

SequentialAgent 依次运行子代理。每个代理都会看到来自前面代理累积的对话历史。

何时使用

• 多步骤流水线，其中输出会进入下一步
• 研究 → 分析 → 总结 工作流
• 数据转换链

完整示例

替换 src/main.rs：

use adk_rust::prelude::*;
use adk_rust::Launcher;
use std::sync::Arc;

#[tokio::main]
async fn main() -> std::result::Result<(), Box<dyn std::error::Error>> {
    dotenvy::dotenv().ok();
    let api_key = std::env::var("GOOGLE_API_KEY")?;
    let model = Arc::new(GeminiModel::new(&api_key, "gemini-2.5-flash")?);

    // Step 1: Research agent gathers information
    let researcher = LlmAgentBuilder::new("researcher")
        .instruction("Research the given topic. List 3-5 key facts or points. \
                     Be factual and concise.")
        .model(model.clone())
        .output_key("research")  // Saves output to state
        .build()?;

    // Step 2: Analyzer agent identifies patterns
    let analyzer = LlmAgentBuilder::new("analyzer")
        .instruction("Based on the research above, identify 2-3 key insights \
                     or patterns. What's the bigger picture?")
        .model(model.clone())
        .output_key("analysis")
        .build()?;

    // Step 3: Summarizer creates final output
    let summarizer = LlmAgentBuilder::new("summarizer")
        .instruction("Create a brief executive summary combining the research \
                     and analysis. Keep it under 100 words.")
        .model(model.clone())
        .build()?;

    // Create the sequential pipeline
    let pipeline = SequentialAgent::new(
        "research_pipeline",
        vec![Arc::new(researcher), Arc::new(analyzer), Arc::new(summarizer)],
    ).with_description("Research → Analyze → Summarize");

    println!("📋 Sequential Pipeline: Research → Analyze → Summarize");
    println!();

    Launcher::new(Arc::new(pipeline)).run().await?;
    Ok(())
}

运行它：

cargo run

示例交互

You: Tell me about Rust programming language

🔄 [researcher] Researching...
Here are key facts about Rust:
1. Systems programming language created at Mozilla in 2010
2. Memory safety without garbage collection via ownership system
3. Zero-cost abstractions and minimal runtime
4. Voted "most loved language" on Stack Overflow for 7 years
5. Used by Firefox, Discord, Dropbox, and Linux kernel

🔄 [analyzer] Analyzing...
Key insights:
1. Rust solves the memory safety vs performance tradeoff
2. Strong developer satisfaction drives rapid adoption
3. Trust from major tech companies validates production-readiness

🔄 [summarizer] Summarizing...
Rust is a systems language that achieves memory safety without garbage 
collection through its ownership model. Created at Mozilla in 2010, it's 
been rated the most loved language for 7 consecutive years. Major companies 
like Discord and Linux kernel adopt it for its zero-cost abstractions 
and performance guarantees.

工作原理

┌─────────────┐    ┌─────────────┐    ┌─────────────┐
│  Researcher │ →  │   Analyzer  │ →  │  Summarizer │
│   (step 1)  │    │   (step 2)  │    │   (step 3)  │
└─────────────┘    └─────────────┘    └─────────────┘
       ↓                  ↓                  ↓
 "Key facts..."    "Insights..."    "Executive summary"

1. 用户消息先发送给第一个代理（Researcher）
2. Researcher 的回复会被添加到历史记录中
3. Analyzer 看到：用户消息 + researcher 回复
4. Summarizer 看到：用户消息 + researcher + analyzer 的回复
5. 当最后一个代理完成时，流水线结束

---

ParallelAgent

ParallelAgent 并发运行所有子代理。每个代理接收相同的输入并独立工作。

这种独立性是被强制保证的，而不只是设计意图：每个子代理都运行在自己的 对话分支（{parent}.{parallel_agent}.{sub_agent}）上，历史读取也 限定在该分支内。子代理能看到导致 分叉的那段对话，但看不到其他兄弟分支产生的内容，因此并发分支不会 相互污染上下文。当你希望子代理之间协同时，请使用 ParallelAgent::with_shared_state()。

分支会一起轮询，因此总耗时大致等于最慢分支的时间， 而不是所有分支时间之和，并且较慢的分支不会拖住 其他分支。设计时值得注意两个后果：

• 事件会交错。 事件按各分支产生的顺序到达，而不是 按子代理分组。使用 event.author（或 event.branch）来标识某个 块属于哪个分支。跨分支的顺序不作保证。
• 一个错误会结束运行，且以确定性方式选定。 每个分支仍会 运行到完成，之后才会抛出一个终止错误。当多个分支失败时， 报告的错误来自声明顺序中最早的 子代理——而不是按实际墙钟时间最先失败的那个，因为并发会使这变成 竞争条件。

过早丢弃事件流会拆除仍在执行中的分支。

何时使用

• 对同一主题进行多视角分析
• 分叉处理（相同输入，不同分析）
• 时间敏感的多任务场景

完整示例

use adk_rust::prelude::*;
use adk_rust::Launcher;
use std::sync::Arc;

#[tokio::main]
async fn main() -> std::result::Result<(), Box<dyn std::error::Error>> {
    dotenvy::dotenv().ok();
    let api_key = std::env::var("GOOGLE_API_KEY")?;
    let model = Arc::new(GeminiModel::new(&api_key, "gemini-2.5-flash")?);

    // Three analysts with DISTINCT personas (important for parallel execution)
    let technical = LlmAgentBuilder::new("technical_analyst")
        .instruction("You are a senior software architect. \
                     FOCUS ONLY ON: code quality, system architecture, scalability, \
                     security vulnerabilities, and tech stack choices. \
                     Start your response with '🔧 TECHNICAL:' and give 2-3 bullet points.")
        .model(model.clone())
        .build()?;

    let business = LlmAgentBuilder::new("business_analyst")
        .instruction("You are a business strategist and MBA graduate. \
                     FOCUS ONLY ON: market opportunity, revenue model, competition, \
                     cost structure, and go-to-market strategy. \
                     Start your response with '💼 BUSINESS:' and give 2-3 bullet points.")
        .model(model.clone())
        .build()?;

    let user_exp = LlmAgentBuilder::new("ux_analyst")
        .instruction("You are a UX researcher and designer. \
                     FOCUS ONLY ON: user journey, accessibility, pain points, \
                     visual design, and user satisfaction metrics. \
                     Start your response with '🎨 UX:' and give 2-3 bullet points.")
        .model(model.clone())
        .build()?;

    // Create parallel agent
    let multi_analyst = ParallelAgent::new(
        "multi_perspective",
        vec![Arc::new(technical), Arc::new(business), Arc::new(user_exp)],
    ).with_description("Technical + Business + UX analysis in parallel");

    println!("⚡ Parallel Analysis: Technical | Business | UX");
    println!("   (All three run simultaneously!)");
    println!();

    Launcher::new(Arc::new(multi_analyst)).run().await?;
    Ok(())
}

💡 提示：使用独特的人设、关注点和回复前缀，让并行代理指令高度区分。这样可以确保每个代理产生独特的输出。

示例交互

You: Evaluate a mobile banking app

🔧 TECHNICAL:
• Requires robust API security: OAuth 2.0, certificate pinning, encrypted storage
• Offline mode with sync requires complex state management and conflict resolution
• Biometric auth integration varies significantly across iOS/Android platforms

💼 BUSINESS:
• Highly competitive market - need unique differentiator (neobanks, traditional banks)
• Revenue model: interchange fees, premium tiers, or lending products cross-sell
• Regulatory compliance costs significant: PCI-DSS, regional banking laws, KYC/AML

🎨 UX:
• Critical: fast task completion - check balance must be < 3 seconds
• Accessibility essential: screen reader support, high contrast mode, large touch targets
• Trust indicators important: security badges, familiar banking patterns

工作原理

                    ┌─────────────────┐
                    │  User Message   │
                    └────────┬────────┘
         ┌───────────────────┼───────────────────┐
         ↓                   ↓                   ↓
  ┌─────────────┐    ┌─────────────┐    ┌─────────────┐
  │  Technical  │    │  Business   │    │     UX      │
  │   Analyst   │    │   Analyst   │    │   Analyst   │
  └──────┬──────┘    └──────┬──────┘    └──────┬──────┘
         ↓                   ↓                   ↓
    (response 1)       (response 2)       (response 3)

所有代理同时启动，并在完成时流式返回结果。

---

LoopAgent

LoopAgent 会重复运行子代理，直到满足退出条件或达到最大迭代次数。

何时使用

• 迭代式改进（草稿 → 评审 → 优化 → 重复）
• 带改进的重试逻辑
• 需要多轮处理的质量门控

ExitLoopTool

要提前退出循环，把 ExitLoopTool 交给某个代理。被调用时，它会向循环发出停止信号。

完整示例

use adk_rust::prelude::*;
use adk_rust::Launcher;
use std::sync::Arc;

#[tokio::main]
async fn main() -> std::result::Result<(), Box<dyn std::error::Error>> {
    dotenvy::dotenv().ok();
    let api_key = std::env::var("GOOGLE_API_KEY")?;
    let model = Arc::new(GeminiModel::new(&api_key, "gemini-2.5-flash")?);

    // Critic agent evaluates content
    let critic = LlmAgentBuilder::new("critic")
        .instruction("Review the content for quality. Score it 1-10 and list \
                     specific improvements needed. Be constructive but critical.")
        .model(model.clone())
        .build()?;

    // Refiner agent improves based on critique
    let refiner = LlmAgentBuilder::new("refiner")
        .instruction("Apply the critique to improve the content. \
                     If the score is 8 or higher, call exit_loop to finish. \
                     Otherwise, provide an improved version.")
        .model(model.clone())
        .tool(Arc::new(ExitLoopTool::new()))  // Can exit the loop
        .build()?;

    // Create inner sequential: critic → refiner
    let critique_refine = SequentialAgent::new(
        "critique_refine_step",
        vec![Arc::new(critic), Arc::new(refiner)],
    );

    // Wrap in loop with max 3 iterations
    let iterative_improver = LoopAgent::new(
        "iterative_improver",
        vec![Arc::new(critique_refine)],
    ).with_max_iterations(3)
     .with_description("Critique-refine loop (max 3 passes)");

    println!("🔄 Iterative Improvement Loop");
    println!("   critic → refiner → repeat (max 3x or until quality >= 8)");
    println!();

    Launcher::new(Arc::new(iterative_improver)).run().await?;
    Ok(())
}

示例交互

You: Write a tagline for a coffee shop

🔄 Iteration 1
[critic] Score: 5/10. "Good coffee here" is too generic. Needs:
- Unique value proposition
- Emotional connection
- Memorable phrasing

[refiner] Improved: "Where every cup tells a story"

🔄 Iteration 2
[critic] Score: 7/10. Better! But could be stronger:
- More action-oriented
- Hint at the experience

[refiner] Improved: "Brew your perfect moment"

🔄 Iteration 3
[critic] Score: 8/10. Strong, action-oriented, experiential.
Minor: could be more distinctive.

[refiner] Score is 8+, quality threshold met!
[exit_loop called]

Final: "Brew your perfect moment"

工作原理

     ┌──────────────────────────────────────────┐
     │              LoopAgent                    │
     │  ┌────────────────────────────────────┐  │
     │  │        SequentialAgent              │  │
     │  │  ┌──────────┐    ┌──────────────┐  │  │
  →  │  │  │  Critic  │ →  │   Refiner    │  │  │  →
     │  │  │ (review) │    │ (improve or  │  │  │
     │  │  └──────────┘    │  exit_loop)  │  │  │
     │  │                  └──────────────┘  │  │
     │  └────────────────────────────────────┘  │
     │         ↑_____________↓                  │
     │         repeat until exit                │
     └──────────────────────────────────────────┘

---

ConditionalAgent（基于规则）

ConditionalAgent 基于一个同步的、基于规则的条件来分支执行。可用于 A/B 测试或基于环境的路由等确定性路由。

ConditionalAgent::new("router", |ctx| ctx.session().state().get("premium")..., premium_agent)
    .with_else(basic_agent)

注意： 对于基于 LLM 的智能路由，请改用 LlmConditionalAgent。

---

LlmConditionalAgent（基于 LLM）

LlmConditionalAgent 使用一个 LLM来分类 用户输入，并路由到合适的子代理。这非常适合需要理解内容的智能路由场景。

何时使用

• 意图分类 - 根据用户在问什么来路由
• 多路由 - 超过 2 个目的地
• 上下文感知路由 - 需要理解，而不仅是关键词

完整示例

use adk_rust::prelude::*;
use adk_rust::Launcher;
use std::sync::Arc;

#[tokio::main]
async fn main() -> std::result::Result<(), Box<dyn std::error::Error>> {
    dotenvy::dotenv().ok();
    let api_key = std::env::var("GOOGLE_API_KEY")?;
    let model = Arc::new(GeminiModel::new(&api_key, "gemini-2.5-flash")?);

    // Create specialist agents
    let tech_agent: Arc<dyn Agent> = Arc::new(
        LlmAgentBuilder::new("tech_expert")
            .instruction("You are a senior software engineer. Be precise and technical.")
            .model(model.clone())
            .build()?
    );

    let general_agent: Arc<dyn Agent> = Arc::new(
        LlmAgentBuilder::new("general_helper")
            .instruction("You are a friendly assistant. Explain simply, use analogies.")
            .model(model.clone())
            .build()?
    );

    let creative_agent: Arc<dyn Agent> = Arc::new(
        LlmAgentBuilder::new("creative_writer")
            .instruction("You are a creative writer. Be imaginative and expressive.")
            .model(model.clone())
            .build()?
    );

    // LLM classifies the query and routes accordingly
    let router = LlmConditionalAgent::builder("smart_router", model.clone())
        .instruction("Classify the user's question as exactly ONE of: \
                     'technical' (coding, debugging, architecture), \
                     'general' (facts, knowledge, how-to), \
                     'creative' (writing, stories, brainstorming). \
                     Respond with ONLY the category name.")
        .route("technical", tech_agent)
        .route("general", general_agent.clone())
        .route("creative", creative_agent)
        .default_route(general_agent)
        .build()?;

    println!("🧠 LLM-Powered Intelligent Router");
    Launcher::new(Arc::new(router)).run().await?;
    Ok(())
}

示例交互

You: How do I fix a borrow error in Rust?
[Routing to: technical]
[Agent: tech_expert]
A borrow error occurs when Rust's ownership rules are violated...

You: What's the capital of France?
[Routing to: general]
[Agent: general_helper]
The capital of France is Paris! It's a beautiful city...

You: Write me a haiku about the moon
[Routing to: creative]
[Agent: creative_writer]
Silver orb above,
Shadows dance on silent waves—
Night whispers secrets.

工作原理

┌─────────────────┐
│  User Message   │
└────────┬────────┘
         ↓
┌─────────────────┐
│   LLM Classifies│  "technical" / "general" / "creative"
│   (smart_router)│
└────────┬────────┘
         ↓
    ┌────┴────┬──────────┐
    ↓         ↓          ↓
┌───────┐ ┌───────┐ ┌─────────┐
│ tech  │ │general│ │creative │
│expert │ │helper │ │ writer  │
└───────┘ └───────┘ └─────────┘

---

组合工作流代理

工作流代理可以嵌套，以实现复杂模式。

顺序 + 并行 + 循环

use adk_rust::prelude::*;
use std::sync::Arc;

// 1. Parallel analysis from multiple perspectives
let parallel_analysis = ParallelAgent::new(
    "multi_analysis",
    vec![Arc::new(tech_analyst), Arc::new(biz_analyst)],
);

// 2. Synthesize the parallel results
let synthesizer = LlmAgentBuilder::new("synthesizer")
    .instruction("Combine all analyses into a unified recommendation.")
    .model(model.clone())
    .build()?;

// 3. Quality loop: critique and refine
let quality_loop = LoopAgent::new(
    "quality_check",
    vec![Arc::new(critic), Arc::new(refiner)],
).with_max_iterations(2);

// Final pipeline: parallel → synthesize → quality loop
let full_pipeline = SequentialAgent::new(
    "full_analysis_pipeline",
    vec![
        Arc::new(parallel_analysis),
        Arc::new(synthesizer),
        Arc::new(quality_loop),
    ],
);

---

跟踪工作流执行

要查看工作流内部发生了什么，请启用 tracing：

use adk_rust::prelude::*;
use adk_rust::{SessionId, UserId};
use adk_rust::runner::{Runner, RunnerConfig};
use adk_rust::futures::StreamExt;
use std::sync::Arc;

// Create pipeline as before...

// Use Runner instead of Launcher for detailed control
let session_service = Arc::new(InMemorySessionService::new());
let runner = Runner::new(RunnerConfig {
    app_name: "workflow_trace".to_string(),
    agent: Arc::new(pipeline),
    session_service: session_service.clone(),
    artifact_service: None,
    memory_service: None,
    run_config: None,
})?;

let session = session_service.create(CreateRequest {
    app_name: "workflow_trace".to_string(),
    user_id: "user".to_string(),
    session_id: None,
    state: Default::default(),
}).await?;

let mut stream = runner.run(
    UserId::new("user")?,
    SessionId::new(session.id())?,
    Content::new("user").with_text("Analyze Rust"),
).await?;

// Process each event to see workflow execution
while let Some(event) = stream.next().await {
    let event = event?;
    
    // Show which agent is responding
    println!("📍 Agent: {}", event.author);
    
    // Show the response content
    if let Some(content) = event.content() {
        for part in &content.parts {
            if let Part::Text { text } = part {
                println!("   {}", text);
            }
        }
    }
    println!();
}

---

API 参考

SequentialAgent

SequentialAgent::new("name", vec![agent1, agent2, agent3])
    .with_description("Optional description")
    .before_callback(callback)  // Called before execution
    .after_callback(callback)   // Called after execution

ParallelAgent

ParallelAgent::new("name", vec![agent1, agent2, agent3])
    .with_description("Optional description")
    .before_callback(callback)
    .after_callback(callback)

如果任何子代理失败，ParallelAgent 会在传播第一个错误之前清空所有剩余的 future，防止资源泄漏。

LoopAgent

LoopAgent::new("name", vec![agent1, agent2])
    .with_max_iterations(5)     // Safety limit (recommended, default: 1000)
    .with_description("Optional description")
    .before_callback(callback)
    .after_callback(callback)

ConditionalAgent

ConditionalAgent::new("name", |ctx| condition_fn, if_agent)
    .with_else(else_agent)      // Optional else branch
    .with_description("Optional description")

ExitLoopTool

// Add to an agent to let it exit a LoopAgent
.tool(Arc::new(ExitLoopTool::new()))

---

上一篇：LlmAgent | 下一篇：多代理系统 →