大语言模型 15 - Manus 超强智能体 开源版本 OpenManus 案例与原理深入解析

写在前面

Manus 是由中国初创公司 Monica.im 于 2025 年 3 月推出的全球首款通用型 AI 智能体（AI Agent），旨在实现“知行合一”，即不仅具备强大的语言理解和推理能力，还能自主执行复杂任务，直接交付完整成果。

什么是 Manus？

“Manus”一词源自拉丁语“Mens et Manus”，意为“心智与双手”，强调其不仅能思考，还能动手执行。与传统 AI 助手（如 ChatGPT）主要提供建议或回答问题不同，Manus 能够独立完成从任务规划到执行的全过程。例如，用户只需输入一句话，Manus 就能自动完成数据分析、代码编写、文档生成等任务，真正实现从想法到成果的闭环。

Manus AI

https://manus.im/

这里放几个用户的案例：
● https://manus.im/share/jr5CKBMJJZDxvU2S5VdZLw?replay=1
● https://manus.im/share/c3onakN6Iajcm1Vt1xAVG7?replay=1
● https://manus.im/share/xFgpHb15vKqfRPWIs3JJPJ?replay=1

应用场景

• 数据分析与可视化：处理上传的数据，生成图表、趋势分析或仪表盘。
• 市场调研：搜索网络信息，整合数据并输出结构化分析结果。
• 旅行规划：根据用户偏好生成详细行程和旅行手册。
• 代码编写与运行：生成代码并测试，提供可执行程序或脚本。
• 文件操作与整理：解压文件、对比文档并生成比较表格。
• 网页生成：根据需求设计并输出展示型网页。
• 多任务并行处理：支持同时处理多个任务，提升工作效率。

案例测试

贪吃蛇

写的提示词如下：

用HTML+CSS+JS 实现一个经典的 贪吃蛇游戏

最终的任务计划总结：

The plan to develop a classic Snake game using HTML, CSS, and JavaScript has been successfully completed. All 10 steps of the plan have been accomplished, achieving a 100% completion rate. The project involved setting up the project structure and basic files, creating and styling the game canvas with HTML and CSS, and initializing game variables with JavaScript. The core functionalities implemented include the snake's movement logic, food generation, collision detection, and game-over conditions. Additionally, features for starting and restarting the game were added. The game underwent testing and debugging to ensure functionality, and the code was optimized with added comments for clarity. The project is now complete with no pending tasks.

实现效果如下：https://wzk.icu/manus/snake/index.html

PingPong

提示词是这样的：

用HTML+CSS+JS 实现一个经典的 PingPong 游戏

计划汇总:

The plan to develop a Classic PingPong Game using HTML, CSS, and JavaScript (ID: plan_1741672648) has been successfully completed. All 9 steps of the plan have been accomplished, resulting in a fully functional game. The project began with setting up the necessary project structure, followed by designing the game layout. The core game logic was implemented using JavaScript, including the mechanics for paddle and ball movement. Collision detection was added to ensure interactions between paddles and walls were accurately represented. A scoring system and game over conditions were established to enhance gameplay. The game underwent thorough testing to ensure functionality and any bugs were addressed. Performance and responsiveness were optimized to provide a smooth user experience. Finally, the game was deployed to a web server or platform, making it accessible for users to play. The project was completed efficiently with no steps left in progress or blocked.

实现结果如下：https://wzk.icu/manus/pingpong/index.html

TODO List

提示词如下：

用HTML+CSS+JS 实现一个经典的todolist 可以添加修改删除

计划汇总如下：

The plan to implement a TODO list using HTML, JavaScript, and CSS has been successfully completed. All nine steps outlined in the plan have been accomplished, achieving a 100% completion rate. The steps included setting up the basic HTML structure, styling the list with CSS, and implementing various JavaScript functionalities such as adding, deleting, editing, and clearing items. Additionally, the application was tested across different browsers to ensure compatibility, and the code was optimized for performance and readability. Finally, the TODO list application was deployed to a web server, marking the completion of the project. The plan was executed efficiently with no steps blocked or left in progress.

效果如下：https://wzk.icu/manus/todolist/index.html

原理分析

基本介绍

OpenManus 采用了一种 类似 Auto-GPT 的思考方式，即智能体不会直接输出答案，而是会：
● 分析任务需求
● 制定解决方案
● 分解任务并执行
● 检查结果并优化
● 返回最终答案

按照我之前的认知来看，是一种：ReACT 的模式。
除了 ReACT 以外，还有一个：CodeACT（Manus团队采访的时候说用的是 CodeACT）：

• https://github.com/xingyaoww/code-act

AutoGPT

● https://github.com/Significant-Gravitas/AutoGPT
● https://agpt.co/

（之前自动生成绘本的项目）在24年的时候，当时使用 AutoGPT做了一些实验，当时GPT-4还非常的昂贵，做了一个给定任务：写小红帽的故事，借助生图工具（DALLE），保持上下文统一风格，最终生成了一组故事。

在 LangChain 中有一个简单案例，可以轻松的实现 ReACT：

# 下面是提示词的模板，通过提示词来激发大模型的 ReACT 行为：
# Answer the following questions as best you can. You have access to the following tools:
#
# {tools}
#
# Use the following format:
#
# Question: the input question you must answer
# Thought: you should always think about what to do
# Action: the action to take, should be one of [{tool_names}]
# Action Input: the input to the action
# Observation: the result of the action
# ... (this Thought/Action/Action Input/Observation can repeat N times)
# Thought: I now know the final answer
# Final Answer: the final answer to the original input question
#
# Begin!
#
# Question: {input}
# Thought:{agent_scratchpad}

# Construct the ReAct agent
# tools = [搜索引擎、绘图、SQL 等等]
agent = create_react_agent(llm, tools, prompt)
# Create an agent executor by passing in the agent and tools
agent_executor = AgentExecutor(agent=agent, tools=tools, verbose=True)

message = agent_executor.invoke({"input": "what is LangChain?"})
print(f"message1: {message}")