传统的串行工作流
def manual_research_pipeline():
     1. 人工文献（耗时2-3小时）
    papers = search_arxiv_by_hand("LLM Agent")
     2. 逐篇阅读摘要（耗时4-5小时）
    summaries = []
    for paper in papers:
        summary = manually_read_and_summarize(paper)
        summaries.append(summary)
     3. 手动整合报告（耗时2小时）
    report = manually_write_report(summaries)
    return report

 ReAct模式伪代码示例
class ReactAgent:
    def run(self, task):
        while not task_completed:
            thought = self.think(current_state)    推理下一步
            action = self.choose_action(thought)   选择行动
            observation = self.execute(action)     执行并观察
            current_state = self.reflect(thought, action, observation)
        return final_result

import json
from typing import List, Dict

class SimpleResearchAssistant:
    """AI实验助手简化实现"""
    
    def __init__(self, llm_api=None):
        self.llm = llm_api or self._mock_llm()
        self.memory = []   记忆存储
    
    def _mock_llm(self):
        """模拟LLM（实际可替换为OpenAI/Claude等API）"""
        return lambda prompt: self._mock_response(prompt)
    
    def _mock_response(self, prompt):
        """模拟响应：解析用户意图并返回规划"""
        if "分析" in prompt:
            return json.dumps({
                "plan": ["search_literature", "extract_keypoints", "generate_summary"],
                "reasoning": "需要先文献，然后提取关键信息，最后生成报告"
            })
        return {"plan": ["default_action"]}
    
    def plan(self, user_task: str) -> List[str]:
        """阶段1：规划 - 将用户任务拆解为步骤"""
        response = self.llm(f"任务：{user_task}，请给出执行步骤")
        steps = json.loads(response) if isinstance(response, str) else response
        self.memory.append({"role": "plan", "task": user_task, "steps": steps})
        return steps["plan"]
    
    def execute(self, step: str, context: Dict = None) -> Dict:
        """阶段2：执行 - 调用具体工具完成子任务"""
         模拟不同工具的执行逻辑
        tools = {
            "search_literature": lambda: {"papers": ["Paper A", "Paper B"], "count": 2},
            "extract_keypoints": lambda: {"keypoints": ["key finding 1", "key finding 2"]},
            "generate_summary": lambda: {"summary": "AI实验助手能够显著提升科研效率..."},
            "default_action": lambda: {"status": "completed"}
        }
        result = tools.get(step, tools["default_action"])()
        self.memory.append({"role": "execute", "step": step, "result": result})
        return result
    
    def run(self, task: str) -> Dict:
        """运行AI实验助手：plan → execute循环"""
        print(f"[AI实验助手] 收到任务：{task}")
        steps = self.plan(task)
        print(f"[规划] 执行步骤：{steps}")
        
        results = {}
        for step in steps:
            print(f"[执行] 正在执行：{step}")
            results[step] = self.execute(step)
        
        return {
            "task": task,
            "steps_completed": steps,
            "results": results,
            "summary": f"AI实验助手已完成{len(steps)}个步骤"
        }

 运行示例
assistant = SimpleResearchAssistant()
output = assistant.run("分析2025年AI Agent领域的论文进展")

print("\n=== 最终输出 ===")
print(json.dumps(output, ensure_ascii=False, indent=2))

 上下文工程示意
class ContextEngine:
    def build_context(self, user_query, memory, tools, external_docs):
        context = {
            "system": "你是一名AI实验助手，请按以下步骤执行...",
            "history": memory.get_relevant_history(user_query),   动态检索
            "tools": tools.get_available_tools(),
            "knowledge": external_docs.search(user_query)         RAG注入
        }
        return self.prioritize(context)   根据token预算压缩

为什么需要？ → AI实验助手是什么？ → Agent vs Workflow → 代码实现 → 底层原理 → 面试考点
  痛点切入        概念定义+类比          关系梳理        实战演练      技术支撑      查漏补缺