跳转至

Agent Frameworks: Assemble Your Team | Agent 框架：组建你的战队

The Mission You are the Commander (Nick Fury). You need to assemble a team of superheroes (Agents) to save the world (Complete a Task). 任务 你是指挥官（尼克·弗瑞）。你需要组建一支超级英雄战队（Agent）来拯救世界（完成任务）。 你应该用哪个框架来管理他们？

1. The Three Frameworks | 三大框架

1.1 LangChain: The Iron Man Lab (钢铁侠实验室)

  • Analogy: You have a garage full of parts (Chains, Prompts, Tools). You build your own suit from scratch.
  • Pros: You can build anything. A flying suit? A tank? A toaster? Yes.
  • Cons: It's complicated. You need to be an engineer like Tony Stark.
  • Best For: Inventors who want total control.

1.2 CrewAI: The Avengers Team (复仇者联盟)

  • Analogy: You have a team of specialists.
    • Hulk: Smash things (Heavy tasks).
    • Black Widow: Spy (Research).
    • Captain America: Leader (Manager).
  • Pros: Everyone knows their role. They work together smoothly.
  • Cons: Hard to make Hulk do spy work (Rigid roles).
  • Best For: Managers who want a reliable workflow.

1.3 AutoGen: The Round Table (圆桌会议)

  • Analogy: You put a bunch of geniuses in a room and let them talk.
    • User Proxy: "I want a snake game."
    • Coder Agent: "Here is the Python code."
    • Reviewer Agent: "Wait, there is a bug. Fix it."
    • Coder Agent: "Fixed. Try again."
  • Pros: They can fix their own mistakes!
  • Cons: Sometimes they argue forever and nothing gets done.
  • Best For: Coders and Problem Solvers.

2. How Agents "Think": The ReAct Loop | Agent 如何“思考”：ReAct 循环

How does an Agent actually do stuff? It uses a magic spell called ReAct (Reason + Act). Agent 到底是怎么做事的？它使用一种叫 ReAct（推理 + 行动）的魔法咒语。

  1. Thought (思考): "The user wants the weather in Tokyo. I don't know it."
  2. Action (行动): "I will use the Google Search tool."
  3. Observation (观察): "Google says it is 25°C and sunny."
  4. Thought (思考): "Now I have the answer."
  5. Final Answer (最终答案): "It is 25°C in Tokyo."
graph TD
    Start[User Request] --> Thought[Thought: What should I do?]
    Thought --> Action[Action: Use Tool]
    Action --> Tool[Tool Output]
    Tool --> Observation[Observation: Read Output]
    Observation --> Thought
    Thought --> Finish[Final Answer]

3. Code Battle: CrewAI Example | 代码对决：CrewAI 示例

Let's build a mini news team. 让我们建立一个迷你新闻团队。

from crewai import Agent, Task, Crew

# 1. The Scout (侦察兵)
# He finds the enemies (news).
researcher = Agent(
    role='Scout',
    goal='Find news about AI',
    backstory='You are a fast scout who reads everything.'
)

# 2. The Bard (吟游诗人)
# He tells the story.
writer = Agent(
    role='Bard',
    goal='Write a legend',
    backstory='You write epic stories about technology.'
)

# 3. The Mission (任务)
task1 = Task(description='Find 3 AI trends', agent=researcher)
task2 = Task(description='Write a blog post', agent=writer)

# 4. Assemble! (集结！)
crew = Crew(agents=[researcher, writer], tasks=[task1, task2])
crew.kickoff()

4. Scientist's Corner | 科学家角落

The Challenge of Planning Humans are great at planning ("I will go to the store, THEN cook dinner"). LLMs are bad at long-term planning. They tend to get distracted. Agentic Workflows (like LangGraph) try to fix this by forcing the model to follow a strict map (State Machine), so it doesn't get lost. 规划的挑战 人类擅长规划（“我会去商店，然后做晚饭”）。 LLM 不擅长长期规划。它们容易分心。 代理工作流（如 LangGraph）试图通过强迫模型遵循严格的地图（状态机）来解决这个问题，这样它就不会迷路。

5. Practice Mission: Design Your Squad | 练习任务：设计你的战队

Objective (目标): Design a multi-agent team to solve a complex problem. 目标：设计一个多智能体团队来解决一个复杂问题。

Scenario (场景): You want to create an "AI Travel Agency". The user says: "Plan a 3-day trip to Kyoto for me. I like temples and sushi." 你想创建一个“AI 旅行社”。 用户说：“帮我计划一个去京都的 3 天旅行。我喜欢寺庙和寿司。”

Task (任务): Define the roles for 3 Agents. 为 3 个 Agent 定义角色。

  1. Agent 1 (The Planner):
    • Role: ?
    • Goal: ?
    • Tools: ? (e.g., Calendar, Map)
  2. Agent 2 (The Researcher):
    • Role: ?
    • Goal: ?
    • Tools: ? (e.g., Google Search, TripAdvisor)
  3. Agent 3 (The Concierge):
    • Role: ?
    • Goal: ?
    • Tools: ? (e.g., Booking.com API)

Workflow (工作流): Draw a simple arrow diagram of who talks to whom. 画一个简单的箭头图，说明谁跟谁说话。 * User -> Agent 1 -> Agent 2 ... ?

Example Answer (参考答案): 1. Agent 1 (Manager): Breaks down the request. "Day 1: Temples. Day 2: Sushi." 2. Agent 2 (Researcher): Finds the best temples (Kinkaku-ji) and sushi places (Sushiro). 3. Agent 3 (Writer): Compiles the itinerary into a nice PDF.

Flow: User -> Manager -> Researcher -> Manager -> Writer -> User.