完整、结构化的复杂 Agent 系统模板
这个模板模拟了现代 LLM Agent 架构，包含：
- Planner（任务规划）
- Memory（短期 + 长期）
- Tool 使用
- Executor（执行）
- Reflection（自我反思）
- Multi-step loop（多轮推理）
- Agent 状态管理
整体结构接近很多 GitHub 项目，例如
- LangChain
- AutoGPT
- MetaGPT
- Microsoft AutoGen
复杂 Agent 系统模板（约200行）
""" Complex Agent System Template Architecture: User -> Planner -> Executor -> Tools -> Memory -> Reflection -> Loop """fromdataclassesimportdataclass,fieldfromtypingimportList,Dict,Any,Callableimportjsonimporttimeimportrandom# ==========================# Data Structures# ==========================@dataclassclassTask:description:strstatus:str="pending"@dataclassclassAgentState:goal:strtasks:List[Task]=field(default_factory=list)history:List[str]=field(default_factory=list)reflections:List[str]=field(default_factory=list)step:int=0# ==========================# Memory System# ==========================classMemory:def__init__(self):self.short_term=[]self.long_term=[]defadd_short(self,message):self.short_term.append(message)defadd_long(self,memory):self.long_term.append(memory)defretrieve(self,query):results=[]forminself.long_term:ifquery.lower()inm.lower():results.append(m)returnresults# ==========================# Tool System# ==========================classToolBox:def__init__(self):self.tools:Dict[str,Callable]={}defregister(self,name,func):self.tools[name]=funcdefrun(self,name,*args,**kwargs):ifnamenotinself.tools:raiseException(f"Tool{name}not found")returnself.tools[name](*args,**kwargs)# Example Toolsdefweb_search(query):returnf"Fake search results about{query}"defcalculator(expr):try:returneval(expr)except:return"error"defsave_file(text):filename=f"result_{int(time.time())}.txt"withopen(filename,"w")asf:f.write(text)returnfilename# ==========================# Planner# ==========================classPlanner:defcreate_plan(self,goal)->List[Task]:# fake planning logicsteps=[f"Research about{goal}",f"Analyze information about{goal}",f"Generate final report about{goal}"]return[Task(s)forsinsteps]# ==========================# Executor# ==========================classExecutor:def__init__(self,toolbox:ToolBox,memory:Memory):self.toolbox=toolbox self.memory=memorydefexecute(self,task:Task):result=""if"Research"intask.description:result=self.toolbox.run("search",task.description)elif"Analyze"intask.description:result=f"Analysis:{random.randint(1,100)}insights discovered"elif"Generate"intask.description:result="Final report generated"else:result="Unknown task"self.memory.add_short(result)returnresult# ==========================# Reflection System# ==========================classReflection:defreflect(self,state:AgentState):iflen(state.history)==0:returnNonelast=state.history[-1]if"error"inlast.lower():reflection="Need to adjust strategy"else:reflection="Task executed successfully"state.reflections.append(reflection)returnreflection# ==========================# Agent Core# ==========================classAgent:def__init__(self):self.memory=Memory()self.toolbox=ToolBox()self.toolbox.register("search",web_search)self.toolbox.register("calc",calculator)self.toolbox.register("save",save_file)self.planner=Planner()self.executor=Executor(self.toolbox,self.memory)self.reflection=Reflection()defrun(self,goal):state=AgentState(goal=goal)# Planningstate.tasks=self.planner.create_plan(goal)print("Plan created:")fortinstate.tasks:print("-",t.description)# Execution Loopwhilestate.step<len(state.tasks):task=state.tasks[state.step]print(f"\nStep{state.step+1}:{task.description}")result=self.executor.execute(task)print("Result:",result)state.history.append(result)task.status="done"reflection=self.reflection.reflect(state)ifreflection:print("Reflection:",reflection)state.step+=1print("\nAgent finished goal")returnstate# ==========================# Example Run# ==========================if__name__=="__main__":agent=Agent()goal="AI Agent architectures"state=agent.run(goal)print("\nFinal reflections:")forrinstate.reflections:print("-",r)这个 Agent 架构包含的关键设计
1 Planner
负责任务分解
Goal ↓ Task1 Task2 Task3类似
- BabyAGI
- AutoGPT
2 Memory
两个层级：
Short-term Long-term真实系统一般用
- Vector DB
- RAG
常见组件：
- FAISS
- Chroma
3 Tool Use
Agent核心能力
Agent -> Tool例子
search() calculator() file save()类似
- LangChain Tool system
4 Reflection
Agent自我改进
result -> critique -> improve类似论文：
- Reflexion Agents
- Self-Refine
5 Agent Loop
经典 Agent 执行循环
while not done: plan act observe reflect类似架构：
- Microsoft AutoGen
- CrewAI
如果把这个模板升级到真实 LLM Agent
需要替换3个地方：
Planner → LLM
plan=llm("break goal into steps")Executor → LLM Tool Call
action=llm("choose tool")Reflection → LLM Critique
reflection=llm("analyze failure")完整真实 Agent 架构
USER │ ▼ Planner (LLM) │ ▼ Task Queue │ ▼ Executor Agent / | \ Tools Memory Web │ │ ▼ ▼ Result -> Reflection │ ▼ Loop