# AG2保持了原AutoGen的核心API
from ag2 import AssistantAgent, UserProxyAgent

assistant = AssistantAgent(
    name="researcher",
    llm_config={"model": "gpt-4o", "api_key": "..."},
    system_message="你是一个研究助手"
)
user = UserProxyAgent(name="user", code_execution_config={"use_docker": True})
user.initiate_chat(assistant, message="分析2026年Agent框架市场格局")

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

class AgentState(TypedDict):
    query: str
    research_result: str
    draft: str
    review_comments: list[str]

def research_node(state: AgentState) -> AgentState:
    # 调用研究Agent
    result = research_agent.invoke(state["query"])
    return {"research_result": result}

def writer_node(state: AgentState) -> AgentState:
    draft = writer_agent.invoke(state["research_result"])
    return {"draft": draft}

def reviewer_node(state: AgentState) -> AgentState:
    comments = reviewer_agent.invoke(state["draft"])
    return {"review_comments": comments}

def should_revise(state: AgentState) -> str:
    if len(state["review_comments"]) > 0:
        return "writer"
    return END

graph = StateGraph(AgentState)
graph.add_node("research", research_node)
graph.add_node("writer", writer_node)
graph.add_node("reviewer", reviewer_node)
graph.add_edge("research", "writer")
graph.add_edge("writer", "reviewer")
graph.add_conditional_edges("reviewer", should_revise)

from crewai import Agent, Task, Crew, Process

researcher = Agent(
    role="市场研究分析师",
    goal="分析2026年Agent框架市场趋势",
    backstory="你是一位有10年经验的科技行业分析师",
    tools=[search_tool, web_scraper],
    llm="gpt-4o"
)

writer = Agent(
    role="技术报告撰写人",
    goal="基于研究数据撰写深度分析报告",
    backstory="你擅长将技术分析转化为商业洞察",
    llm="claude-sonnet-4-6"
)

reviewer = Agent(
    role="质量审核员",
    goal="确保报告准确性和完整性",
    backstory="你对数据准确性和逻辑一致性有极高要求",
    llm="gpt-4o"
)

research_task = Task(description="调研六大Agent框架", agent=researcher)
write_task = Task(description="撰写对比报告", agent=writer)
review_task = Task(description="审核报告质量", agent=reviewer)

crew = Crew(
    agents=[researcher, writer, reviewer],
    tasks=[research_task, write_task, review_task],
    process=Process.sequential
)

result = crew.kickoff()

from ag2 import AssistantAgent, GroupChat, GroupChatManager

analyst = AssistantAgent(
    name="data_analyst",
    system_message="分析数据并生成洞察",
    llm_config={"config_list": [{"model": "gpt-4o"}]}
)

critic = AssistantAgent(
    name="critic",
    system_message="批判性地审视分析结果，指出潜在偏差",
    llm_config={"config_list": [{"model": "claude-sonnet-4-6"}]}
)

summarizer = AssistantAgent(
    name="summarizer",
    system_message="将讨论总结为可执行的建议",
    llm_config={"config_list": [{"model": "gpt-4o"}]}
)

group_chat = GroupChat(
    agents=[analyst, critic, summarizer],
    messages=[],
    max_round=10
)

manager = GroupChatManager(groupchat=group_chat)
manager.initiate_chat(
    analyst,
    message="分析2026年Q1 AI Agent市场的关键变化"
)

from agents import Agent, Runner, handoff, guardrail

triage_agent = Agent(
    name="triage",
    instructions="根据用户问题判断应该转交给哪个专家Agent",
    handoffs=[
        handhandoff(target=research_agent, description="研究分析类问题"),
        handoff(target=coding_agent, description="编程开发类问题"),
        handoff(target=creative_agent, description="创意写作类问题")
    ]
)

@guardrail
def check_output_quality(ctx, agent, output):
    if output.confidence < 0.7:
        return {"blocked": True, "reason": "置信度过低"}
    return {"blocked": False}

result = Runner.run_sync(
    starting_agent=triage_agent,
    messages=[{"role": "user", "content": "帮我分析LangGraph和CrewAI的优劣"}],
    guardrails=[check_output_quality]
)

from claude_agent_sdk import Agent, Tool

@Tool
def search_web(query: str) -> str:
    """搜索网络获取最新信息"""
    return web_search(query)

@Tool
def read_file(path: str) -> str:
    """读取本地文件内容"""
    with open(path) as f:
        return f.read()

agent = Agent(
    model="claude-sonnet-4-6",
    tools=[search_web, read_file],
    system_prompt="你是一个研究助手，善于搜索和整合信息"
)

result = agent.run("分析2026年AI Agent市场的最新动态")

from google.adk import Agent, Transfer, Session

research_agent = Agent(
    name="researcher",
    model="gemini-2.5-pro",
    instruction="执行深度研究分析",
    tools=[google_search, scholar_search]
)

writing_agent = Agent(
    name="writer",
    model="gemini-2.5-pro",
    instruction="将研究结果转化为报告"
)

root_agent = Agent(
    name="coordinator",
    model="gemini-2.5-pro",
    instruction="协调研究者和撰写者的工作",
    transfers=[
        Transfer(target=research_agent, description="研究任务"),
        Transfer(target=writing_agent, description="撰写任务")
    ]
)

session = Session(agent=root_agent)
result = session.run("分析2026年Agent框架竞争格局")

graph TB
    subgraph "2026 Agent协议双层架构"
        subgraph "工具层 - MCP"
            M1[数据库] --> MCP[MCP Server]
            M2[API服务] --> MCP
            M3[文件系统] --> MCP
            M4[代码执行器] --> MCP
            MCP --> Agent_A[Agent A]
            MCP --> Agent_B[Agent B]
        end

        subgraph "协调层 - A2A"
            Agent_A <-->|A2A协议| Agent_B
            Agent_B <-->|A2A协议| Agent_C[Agent C<br/>跨组织]
        end
    end

    style MCP fill:#4CAF50,color:#fff
    style A2A fill:#2196F3,color:#fff

graph TD
    START[选择Agent框架] --> Q1{核心需求是什么?}

    Q1 -->|复杂工作流<br/>精确控制| LANG[LangGraph]
    Q1 -->|快速原型<br/>团队协作| CREW[CrewAI]
    Q1 -->|多Agent对话<br/>辩论推理| AG2[AG2]
    Q1 -->|OpenAI生态<br/>简单轻量| OPENAI[OpenAI Agents SDK]
    Q1 -->|Claude生态<br/>MCP工具链| CLAUDE[Claude Agent SDK]
    Q1 -->|GCP部署<br/>A2A通信| GOOGLE[Google ADK]

    LANG --> L1{需要可观测性?}
    L1 -->|是| L2[LangGraph + LangSmith<br/>生产首选]
    L1 -->|否| L3[LangGraph OSS<br/>成本敏感型]

    CREW --> C1{团队规模?}
    C1 -->|<5人| C2[CrewAI OSS<br/>快速迭代]
    C1 -->|>5人/企业| C3[CrewAI Enterprise<br/>可视化编排]

    AG2 --> A1{AutoGen迁移?}
    A1 -->|是| A2[AG2 - API兼容<br/>迁移成本最低]
    A1 -->|否| A3[考虑其他框架<br/>AG2社区风险]

    style LANG fill:#FF6B35,color:#fff
    style CREW fill:#7B68EE,color:#fff
    style AG2 fill:#20B2AA,color:#fff
    style OPENAI fill:#10A37F,color:#fff
    style CLAUDE fill:#D97706,color:#fff
    style GOOGLE fill:#4285F4,color:#fff