graph TB
    subgraph L1["模型原生层 — 零税·单模型深度优化"]
        OAI["OpenAI Agents SDK"]
        CLA["Claude Agent SDK"]
    end

    subgraph L2["编排层 — 跨模型·复杂工作流"]
        LG["LangGraph"]
        ADK["Google ADK 2.0"]
        MAF["MS Agent Framework"]
    end

    subgraph L3["协作层 — 角色化·快速原型"]
        CREW["CrewAI"]
    end

    subgraph L4["原子层 — 极简·工具密集"]
        SMOL["SmolAgents"]
    end

    L1 -->|"需要复杂编排"| L2
    L2 -->|"需要角色抽象"| L3
    L2 -->|"需要轻量工具"| L4

    style OAI fill:#10a37f,color:#fff
    style CLA fill:#d97706,color:#fff
    style LG fill:#1a1a2e,color:#fff
    style ADK fill:#4285f4,color:#fff
    style MAF fill:#00bcf2,color:#fff
    style CREW fill:#ff6b6b,color:#fff
    style SMOL fill:#ffdd57,color:#333

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

class AgentState(TypedDict):
    messages: Annotated[list, operator.add]
    revision_count: int
    approved: bool

def researcher(state):
    return {"messages": ["Research findings..."]}

def writer(state):
    return {"messages": ["Draft v1..."], "revision_count": 0}

def reviewer(state):
    approved = state["revision_count"] >= 2  # 最多修改2轮
    return {"approved": approved, "revision_count": state["revision_count"] + 1}

def route_review(state):
    return END if state["approved"] else "writer"

graph = StateGraph(AgentState)
graph.add_node("research", researcher)
graph.add_node("write", writer)
graph.add_node("review", reviewer)
graph.add_edge("research", "write")
graph.add_edge("write", "review")
graph.add_conditional_edges("review", route_review)
app = graph.compile()

from crewai import Agent, Task, Crew, Process

researcher = Agent(
    role="Senior Research Analyst",
    goal="Uncover cutting-edge AI developments",
    backstory="You are a seasoned analyst at a top tech research firm.",
    tools=[search_tool],
)

writer = Agent(
    role="Tech Content Strategist",
    goal="Craft compelling articles",
    backstory="Former tech journalist, now crafting viral content.",
)

research = Task(
    description="Research latest AI agent frameworks in 2026",
    expected_output="Structured report with key findings",
    agent=researcher,
)

writing = Task(
    description="Write a comprehensive article based on research",
    expected_output="2000-word article with citations",
    agent=writer,
)

crew = Crew(
    agents=[researcher, writer],
    tasks=[research, writing],
    process=Process.sequential,
)
result = crew.kickoff()

from google.adk import Agent, Workflow
from google.adk.tools import MCPTool

researcher = Agent(
    name="researcher",
    model="gemini-2.5-pro",
    instruction="Research AI agent frameworks",
    tools=[MCPTool("web-search-server")],
)

writer = Agent(
    name="writer",
    model="gemini-2.5-flash",  # 写作用便宜模型
    instruction="Write article from research notes",
)

workflow = Workflow()
workflow.add_agent(researcher)
workflow.add_agent(writer)
workflow.add_edge("researcher", "writer")
# 支持条件分支
workflow.add_conditional_edge("writer", lambda s: "researcher" if s.needs_revision else END)

result = await workflow.run(input="Write about AI agents in 2026")

// .NET路径
var analyst = new ChatAgent("analyst")
    .WithPlugin(new WebSearchPlugin())
    .WithPlugin(new DataAnalysisPlugin());

var writer = new ChatAgent("writer")
    .WithPlugin(new DocumentPlugin());

var orchestration = new SequentialOrchestration()
    .AddAgent(analyst)
    .AddAgent(writer);

var result = await orchestration.InvokeAsync(
    "Analyze Q1 sales data and write executive summary"
);

# Python路径
from microsoft.agentframework import ChatAgent, SequentialOrchestration

analyst = ChatAgent("analyst", plugins=[search_plugin, data_plugin])
writer = ChatAgent("writer", plugins=[doc_plugin])

orchestration = SequentialOrchestration([analyst, writer])
result = await orchestration.invoke("Analyze Q1 sales and write summary")

from smolagents import CodeAgent, DuckDuckGoSearchTool, HfApiModel

agent = CodeAgent(
    tools=[DuckDuckGoSearchTool()],
    model=HfApiModel("Qwen/Qwen2.5-72B-Instruct"),
    max_steps=6,
)

# 一行运行
result = agent.run(
    "Search for AI agent framework benchmarks 2026, "
    "compare top 3 and create a summary table"
)

# SmolAgents内部生成：
search_results = web_search("AI agent framework benchmarks 2026")
summary_table = []
for framework in top_3:
    summary_table.append({
        "name": framework.name,
        "stars": framework.github_stars,
        "token_overhead": framework.token_overhead
    })
print(summary_table)

from openai import Agent, Runner
from openai.types.agent import handoff

troubleshooter = Agent(
    name="troubleshooter",
    instructions="Help debug code issues. Escalate to architect if complex.",
    tools=[code_search_tool, log_tool],
)

architect = Agent(
    name="architect",
    instructions="Design system-level solutions for complex problems.",
    tools=[design_tool, review_tool],
    handoffs=[handoff(troubleshooter)],  # 可回退给排查Agent
)

result = Runner.run(
    architect,
    "Our microservice has a memory leak in production",
    max_turns=15,
)

xychart-beta
    title "六栈框架税率对比（1000次推理均值）"
    x-axis ["OpenAI SDK", "SmolAgents", "ADK 2.0", "LangGraph", "MS AF", "CrewAI"]
    y-axis "框架税率 (%)" 0 --> 40
    bar [4, 6, 9, 13.5, 16.5, 31.5]

graph LR
    subgraph "单步 → 10步衰减"
        direction TB
        OAI["OpenAI SDK: 99.6% → 95.8% (-3.8pp)"]
        ADK["Google ADK: 99.5% → 94.1% (-5.4pp)"]
        LG["LangGraph: 99.2% → 91.7% (-7.5pp)"]
        SMOL["SmolAgents: 99.0% → 90.3% (-8.7pp)"]
        MAF["MS AF: 98.8% → 89.5% (-9.3pp)"]
        CREW["CrewAI: 98.5% → 84.2% (-14.3pp)"]
    end

graph TD
    START["选择Agent框架"] --> Q1{"是否已绑定特定模型厂?"}
    Q1 -->|"是, OpenAI"| OAI["OpenAI Agents SDK"]
    Q1 -->|"是, Google"| ADK["Google ADK 2.0"]
    Q1 -->|"是, Microsoft"| MAF["MS Agent Framework"]
    Q1 -->|"否, 需要多模型"| Q2{"任务复杂度?"}

    Q2 -->|"单Agent·工具密集"| SMOL["SmolAgents"]
    Q2 -->|"多Agent·简单流程"| CREW["CrewAI"]
    Q2 -->|"多Agent·复杂流程"| Q3{"成本敏感度?"}

    Q3 -->|"高"| LG1["LangGraph (最低Token开销的编排层)"]
    Q3 -->|"中"| ADK2["Google ADK 2.0 (A2A原生)"]
    Q3 -->|"低"| MAF2["MS Agent Framework (企业集成)"]

    style OAI fill:#10a37f,color:#fff
    style ADK fill:#4285f4,color:#fff
    style MAF fill:#00bcf2,color:#fff
    style SMOL fill:#ffdd57,color:#333
    style CREW fill:#ff6b6b,color:#fff
    style LG1 fill:#1a1a2e,color:#fff
    style ADK2 fill:#4285f4,color:#fff
    style MAF2 fill:#00bcf2,color:#fff