协议原生Agent栈：六向突围实测

from langgraph.graph import StateGraph, END
from typing import TypedDict

class AgentState(TypedDict):
    messages: list
    next_agent: str
    iteration: int

def researcher(state: AgentState) -> AgentState:
    # 调用LLM进行研究
    result = llm.invoke(state["messages"])
    state["messages"].append(result)
    state["next_agent"] = "writer" if result.confidence > 0.8 else "reviewer"
    return state

graph = StateGraph(AgentState)
graph.add_node("researcher", researcher)
graph.add_node("writer", writer_node)
graph.add_node("reviewer", reviewer_node)
graph.add_conditional_edges("researcher", lambda s: s["next_agent"])
graph.add_edge("writer", END)
graph.add_edge("reviewer", "researcher")  # 循环回退

app = graph.compile()

from crewai import Agent, Task, Crew, Process

researcher = Agent(
    role="Senior Research Analyst",
    goal="Discover and analyze emerging AI trends",
    backstory="You are a veteran analyst at a top-tier research firm",
    tools=[search_tool, scrape_tool],
    llm="gpt-4o",
    max_iter=5,
    verbose=True
)

writer = Agent(
    role="Tech Content Strategist",
    goal="Transform analysis into compelling articles",
    backstory="Former tech journalist with 10 years of experience",
    tools=[write_tool],
)

research_task = Task(
    description="Analyze top 5 AI agent frameworks in 2026",
    expected_output="Structured analysis with comparison matrix",
    agent=researcher
)

write_task = Task(
    description="Write a 2000-word analysis article",
    expected_output="Publication-ready markdown article",
    agent=writer
)

crew = Crew(
    agents=[researcher, writer],
    tasks=[research_task, write_task],
    process=Process.sequential,
    memory=True,  # 2026新增：跨Agent共享记忆
)

result = crew.kickoff()

import autogen

# AG2的可组合Agent模式
researcher = autogen.AssistantAgent(
    name="researcher",
    system_message="You are a research specialist.",
    llm_config={"model": "gpt-4o", "cache_seed": 42},
    max_consecutive_auto_reply=5,
)

# 人工代理节点
user_proxy = autogen.UserProxyAgent(
    name="user_proxy",
    human_input_mode="TERMINATE",
    code_execution_config={
        "work_dir": "workspace",
        "use_docker": True,  # 沙箱执行
    },
)

# GroupChat：多Agent圆桌讨论
groupchat = autogen.GroupChat(
    agents=[researcher, coder, reviewer, user_proxy],
    messages=[],
    max_round=20,
    speaker_selection_method="round_robin",  # 2026新增：多种发言选择策略
)

manager = autogen.GroupChatManager(groupchat=groupchat)
user_proxy.initiate_chat(manager, message="Build a web scraper for AI news")

from openai import Agent, Runner, handoff
from openai.tools import function_tool

@function_tool
def search_web(query: str) -> str:
    """Search the web for information."""
    return web_search_api(query)

@function_tool
def analyze_data(data_url: str) -> str:
    """Analyze data from a URL."""
    return data_analysis_pipeline(data_url)

researcher = Agent(
    name="Research Specialist",
    instructions="Research topics thoroughly using web search.",
    tools=[search_web],
)

analyst = Agent(
    name="Data Analyst",
    instructions="Analyze data and provide insights.",
    tools=[analyze_data],
    handoffs=[handoff(researcher, "Need more research on this topic")],
)

# 沙箱执行 - 2026年关键特性
result = await Runner.run(
    analyst,
    messages=[{"role": "user", "content": "Analyze AI market trends"}],
    context={"sandbox": True, "max_turns": 10},
)

import { Agent, Tool, Hook } from '@anthropic-ai/claude-agent-sdk';

// 自定义MCP工具
const databaseTool = Tool.fromMCPServer({
  serverUrl: 'http://localhost:3001',
  toolName: 'query_customer_db',
});

// 安全Hook：过滤敏感信息
const piiFilter = Hook.beforeToolCall(async (context) => {
  const pii = detectPII(context.toolInput);
  if (pii) {
    return { blocked: true, reason: `PII detected: ${pii.type}` };
  }
  return { allowed: true };
});

// 成本控制Hook
const budgetGuard = Hook.afterLLMResponse(async (context) => {
  const cost = context.tokenUsage.total * 0.003; // Claude Sonnet定价
  if (context.sessionCost + cost > 5.0) { // $5会话预算
    return { action: 'terminate', reason: 'Budget exceeded' };
  }
});

const agent = new Agent({
  model: 'claude-sonnet-4-6',
  instructions: 'You are a customer service agent.',
  tools: [databaseTool],
  hooks: [piiFilter, budgetGuard],
  maxTurns: 25,
});

const result = await agent.run('Help me check my order status');

from google.adk import Agent, Workflow, GraphBuilder
from google.adk.a2a import A2AClient

# A2A原生Agent定义
research_agent = Agent(
    name="researcher",
    model="gemini-2.5-pro",
    instruction="Research and summarize AI trends",
    tools=[A2AClient("http://search-service:8080")],  # A2A工具
)

writer_agent = Agent(
    name="writer", 
    model="gemini-2.5-flash",  # 写作用轻量模型
    instruction="Write engaging analysis articles",
)

# 图工作流
graph = GraphBuilder() \
    .add_node("research", research_agent) \
    .add_node("draft", writer_agent) \
    .add_node("review", review_agent) \
    .add_edge("research", "draft") \
    .add_conditional_edge("review", 
        lambda state: "draft" if state.needs_revision else END) \
    .build()

workflow = Workflow(graph=graph, model_routing="auto")
result = await workflow.run("Analyze AI agent frameworks")

graph TB
    subgraph "应用层"
        APP[业务应用]
        UI[用户界面]
    end

    subgraph "编排层 — 六栈分野"
        direction LR
        LG[LangGraph<br/>有向图编排]
        CA[CrewAI<br/>角色编排]
        AG2[AG2<br/>对话式编排]
        OAI[OpenAI SDK<br/>Handoff编排]
        CLA[Claude SDK<br/>工具+Hooks编排]
        ADK[ADK 2.0<br/>图+A2A编排]
    end

    subgraph "协议层 — 统一通信"
        MCP[MCP<br/>工具连接协议<br/>Anthropic 2024.11]
        A2A[A2A<br/>Agent间协议<br/>Google 2025.04]
        ACP[ACP<br/>Agent通信协议<br/>Microsoft 2025]
    end

    subgraph "基础设施层"
        LM[大语言模型]
        MEM[向量存储/记忆]
        SANDBOX[沙箱运行时]
        OBS[可观测性<br/>LangSmith/Phoenix]
    end

    APP --> LG & CA & AG2 & OAI & CLA & ADK
    UI --> APP

    LG -.适配.-> MCP
    CA -.适配.-> MCP
    AG2 -.原生.-> ACP
    OAI -.适配.-> MCP
    CLA -.原生.-> MCP
    ADK -.原生.-> A2A

    MCP --> LM & MEM & SANDBOX
    A2A --> LM & MEM
    ACP --> LM & MEM

    LG & CA & ADK --> OBS

    style MCP fill:#e1f5fe
    style A2A fill:#e8f5e9
    style ACP fill:#fff3e0
    style CLA fill:#fce4ec
    style ADK fill:#e8f5e9
    style LG fill:#f3e5f5

graph LR
    subgraph "协议原生"
        CLA_MCP[Claude Agent SDK] -->|原生| MCP
        ADK_A2A[Google ADK 2.0] -->|原生| A2A
        AG2_ACP[AG2] -->|原生| ACP
    end

    subgraph "协议适配"
        LG_MCP[LangGraph] -->|适配器| MCP
        CA_MCP[CrewAI] -->|适配器| MCP
        OAI_MCP[OpenAI SDK] -->|适配器| MCP
    end

    subgraph "跨协议桥"
        MCP <-->|双向桥| A2A
        MCP <-->|双向桥| ACP
        A2A <-->|双向桥| ACP
    end

    style CLA_MCP fill:#c8e6c9
    style ADK_A2A fill:#c8e6c9
    style AG2_ACP fill:#c8e6c9
    style LG_MCP fill:#fff9c4
    style CA_MCP fill:#fff9c4
    style OAI_MCP fill:#fff9c4

graph TD
    START[选择Agent框架] --> Q1{是否绑定特定云平台?}
    Q1 -->|Azure| AZ[AG2 + Semantic Kernel]
    Q1 -->|GCP| GC[Google ADK 2.0]
    Q1 -->|AWS| Q2{...}
    Q1 -->|无/多云| Q3{主要使用哪个模型?}

    Q3 -->|GPT系列| Q4{团队规模?}
    Q3 -->|Claude系列| CL[Claude Agent SDK]
    Q3 -->|Gemini系列| ADK[Google ADK 2.0]
    Q3 -->|混合/可切换| Q5{需要多Agent编排?}

    Q4 -->|1-3人| OAI[OpenAI Agents SDK]
    Q4 -->|4+人| Q5

    Q5 -->|简单串行| CA2[CrewAI]
    Q5 -->|复杂图流程| LG2[LangGraph]
    Q5 -->|企业级| LG3[LangGraph + LangSmith]

    AZ --> END2[开始开发]
    GC --> END2
    CL --> END2
    ADK --> END2
    OAI --> END2
    CA2 --> END2
    LG2 --> END2
    LG3 --> END2

    style START fill:#e3f2fd
    style END2 fill:#e8f5e9