from langgraph.graph import StateGraph, END
from langgraph.checkpoint.postgres import PostgresSaver

# 定义状态
class AgentState(TypedDict):
    messages: list[BaseMessage]
    next_action: str

# 构建图
graph = StateGraph(AgentState)
graph.add_node("reasoner", reasoner_node)
graph.add_node("tool_executor", tool_node)
graph.add_node("evaluator", evaluator_node)

# 条件边：根据评估结果决定走向
graph.add_conditional_edges("evaluator", route_decision, {
    "continue": "reasoner",
    "finish": END,
    "escalate": "human_review"
})

# 编译并启用检查点
checkpointer = PostgresSaver.from_conn_string("postgresql://...")
app = graph.compile(
    checkpointer=checkpointer,
    interrupt_before=["human_review"]  # 人工审批断点
)

from agents import Agent, Runner, handoff

# 定义专业Agent
triage_agent = Agent(
    name="Triage Agent",
    instructions="分析用户请求并路由到合适的专家",
    handoffs=[
        handoff(research_agent, description="需要深度研究的问题"),
        handoff(code_agent, description="代码编写和调试任务"),
        handoff(data_agent, description="数据分析和可视化"),
    ]
)

research_agent = Agent(
    name="Research Agent",
    instructions="你是研究专家，使用搜索工具回答问题",
    tools=[web_search, academic_search],
    handoffs=[handoff(triage_agent, description="需要转交其他专家")]  # 回传
)

# 运行Agent循环
result = await Runner.run(triage_agent, "分析2026年AI Agent框架趋势")

from pydantic_ai import Agent, RunContext
from pydantic import BaseModel

class SearchResult(BaseModel):
    title: str
    url: str
    relevance: float

# 类型注解自动推断工具签名
research_agent = Agent(
    'openai:gpt-4o',
    result_type=SearchResult,  # 结构化输出自动验证
    deps_type=DatabaseConnection,  # 依赖注入
)

@research_agent.tool
async def search_web(ctx: RunContext[DatabaseConnection], query: str) -> list[str]:
    """搜索工具——参数和返回类型自动生成JSON Schema"""
    results = await ctx.deps.search(query)
    return [r.url for r in results]

# 运行并自动验证结果
result = await research_agent.run("最新Agent框架")
assert isinstance(result.data, SearchResult)  # 类型保证

from crewai import Agent, Task, Crew, Process

# 定义角色Agent
researcher = Agent(
    role="高级研究分析师",
    goal="发现AI Agent领域的突破性技术和趋势",
    backstory="你曾在麦肯锡担任技术分析师10年，擅长从噪音中提取信号",
    tools=[search_tool, scraper_tool],
    verbose=True
)

writer = Agent(
    role="技术内容总监",
    goal="将复杂技术概念转化为引人入胜的深度文章",
    backstory="你是Wired前主编，擅长技术叙事",
    verbose=True
)

# 定义任务链
research_task = Task(
    description="研究2026年最有前景的5个Agent框架",
    agent=researcher,
    expected_output="包含技术细节和对比的结构化报告"
)

write_task = Task(
    description="基于研究报告撰写5000字深度分析文章",
    agent=writer,
    expected_output="Markdown格式的完整文章"
)

# 组建团队并执行
crew = Crew(
    agents=[researcher, writer],
    tasks=[research_task, write_task],
    process=Process.sequential  # 顺序执行
)
result = crew.kickoff()

from microsoft.agent_framework import Agent, AgentRuntime, Workflow

# 定义Agent工作流
workflow = Workflow()

research_agent = Agent(
    name="ResearchAgent",
    instructions="使用Bing搜索收集信息",
    tools=[bing_search_tool],
    model="azure:gpt-4o"
)

analyst_agent = Agent(
    name="AnalystAgent",
    instructions="分析研究结果并生成洞察",
    tools=[chart_generator_tool],
)

# 编排工作流
workflow.add_edge(research_agent, analyst_agent)
workflow.add_edge(analyst_agent, output_handler)

# 运行时
runtime = AgentRuntime(
    workflow=workflow,
    persistence=PostgresPersistence(),
    tracing=AzureMonitorTracing(),  # Azure Monitor集成
)
result = await runtime.execute("分析半导体行业趋势")

from google.adk import Agent, LlmAgent, SequentialAgent
from google.adk.tools import google_search

# 定义Agent
coordinator = Agent(
    name="coordinator",
    model="gemini-2.5-pro",
    instruction="协调多个专家Agent完成用户任务",
)

researcher = LlmAgent(
    name="researcher",
    model="gemini-2.5-flash",
    instruction="深度研究指定主题",
    tools=[google_search],
)

# Graph编排
from google.adk import GraphAgent

graph = GraphAgent(root=coordinator)
graph.add_edge(coordinator, researcher)
graph.add_edge(researcher, writer)
graph.add_edge(writer, coordinator)  # 循环

result = await graph.run("分析2026年6G标准化进展")

from strands import Agent, Tool

# 模型驱动：极简定义
@Tool
def search_database(query: str) -> dict:
    """搜索产品数据库"""
    return db.query(query)

agent = Agent(
    tools=[search_database],
    system_prompt="你是产品查询助手"
)

# 一行运行，框架自动编排工具调用
result = agent("查找价格低于100元的智能手表")

import ag2

# 定义可对话Agent
analyst = ag2.AssistantAgent(
    name="analyst",
    system_message="你是数据分析师，擅长发现数据中的模式",
    llm_config={"model": "gpt-4o"}
)

critic = ag2.AssistantAgent(
    name="critic",
    system_message="你是严格的质量审查员，挑战每一个结论",
    llm_config={"model": "gpt-4o"}
)

# GroupChat多Agent对话
group_chat = ag2.GroupChat(
    agents=[analyst, critic, moderator],
    messages=[],
    max_round=10
)

manager = ag2.GroupChatManager(groupchat=group_chat)
analyst.initiate_chat(manager, message="分析这份销售数据的异常模式")

graph TB
    subgraph "2026 Agent框架架构全景"
        direction TB

        subgraph "编排层 Orchestration"
            LG[LangGraph<br/>状态图编排]
            OA[OpenAI Agents SDK<br/>Handoff原语]
            CR[CrewAI<br/>角色扮演]
            GF[Google ADK 2.0<br/>Graph Workflow]
        end

        subgraph "运行时层 Runtime"
            SK[Microsoft MAF<br/>Actor Model]
            ST[Strands Agents<br/>模型驱动]
            AG[AG2<br/>分布式Actor]
        end

        subgraph "类型安全层 Type Safety"
            PA[PydanticAI<br/>类型注解验证]
        end

        subgraph "协议层 Protocol"
            MCP[MCP<br/>工具调用协议]
            A2A[A2A<br/>Agent间通信]
            ACP[ACP<br/>能力协议]
        end

        subgraph "模型层 Models"
            GPT[GPT-4o/o3]
            GEM[Gemini 2.5]
            CLA[Claude Opus 4]
            BED[Bedrock Models]
        end
    end

    LG --> MCP
    OA --> MCP
    CR --> MCP
    GF --> A2A
    GF --> MCP
    SK --> MCP
    SK --> A2A
    ST --> A2A
    ST --> MCP
    AG --> MCP
    PA --> MCP

    LG --> GPT
    LG --> GEM
    LG --> CLA
    OA --> GPT
    PA --> GPT
    PA --> GEM
    PA --> CLA
    CR --> GPT
    GF --> GEM
    SK --> GPT
    ST --> BED

flowchart TD
    START[开始选型] --> Q1{主要语言?}
    Q1 -->|.NET| MAF[Microsoft Agent Framework]
    Q1 -->|Python| Q2{需要复杂工作流编排?}
    Q1 -->|TypeScript| Q3{云平台?}

    Q2 -->|是,有循环/分支| Q4{厂商锁定容忍度?}
    Q2 -->|否,简单Agent| Q5{团队技术背景?}

    Q4 -->|不介意Google锁定| ADK[Google ADK 2.0]
    Q4 -->|需要模型无关| LG[LangGraph]
    Q4 -->|接受OpenAI绑定| OAS[OpenAI Agents SDK]

    Q5 -->|Pydantic/FastAPI用户| PAI[PydanticAI]
    Q5 -->|需要低门槛多Agent| CREW[CrewAI]
    Q5 -->|AWS用户| STR[Strands Agents]
    Q5 -->|需要分布式| AG2[AG2]

    Q3 -->|AWS| STR
    Q3 -->|GCP| ADK
    Q3 -->|通用| OAS