Unpacking Manus AI Agent: Architecture, Prompt Design, and a Simple DIY Clone

What is Manus?

Manus, released by the Chinese startup Monica, is marketed as the world’s first general‑purpose AI agent capable of planning and autonomously executing complete tasks such as writing reports or generating spreadsheets.

Its name, Latin for “hand”, reflects the idea that knowledge should be turned into action, moving beyond simple information provision.

Product Design

Manus offers a chat‑style input box with two modes:

Standard : Uses non‑reasoning models (e.g., Qwen2.5‑Max, DeepSeek‑V3, GPT‑4.5) but runs slower because many tools are invoked.

High‑Investment : Uses reasoning models (e.g., Qwen‑32B, DeepSeek‑R1, OpenAI o1) which do not expose the reasoning process and consume more tokens.

The interface displays a large‑model output area on the left, a virtual‑computer view in the upper‑right, and a task‑progress panel in the lower‑right.

Technical Design

The visible autonomous execution process can be illustrated with a real‑world example of diagnosing an Alibaba Cloud email domain’s DNS records.

1. Task Planning

Manus first decomposes the user request into coarse‑grained steps that form a global plan, allowing the user to see overall progress.

2. Task Execution

During execution, each coarse step is further broken into fine‑grained sub‑steps in an incremental fashion. When a command needs to run, Manus spawns a remote sandbox VM where all subsequent commands execute, preserving state across the session.

3. Task Reflection

If a command fails (e.g., missing environment or illegal syntax), Manus adjusts the plan, re‑executes, or selects a different command, following the CodeAct principle of self‑debugging.

4. Intermediate Files

TODO List

After each completed sub‑task, Manus updates a todo.md file, creating it if absent and marking finished items with ✅.

Process Files

When intermediate results need to be persisted, Manus writes them to markdown files (e.g., .md) for later reference.

5. Final Output

Once all planned steps finish, Manus assembles the final answer, summarizing the solution and attaching any generated files.

Underlying Design Ideas

Because the official Manus code is closed‑source, the analysis combines observed execution traces, the open‑source OpenManus project, and publicly shared prompts to infer the architecture.

OpenManus

Agent Loop Diagram

OpenManus follows a classic ReAct loop: observe → think → act → observe, implemented with a Step() function that iterates until the task is finished.

Prompt Design

SYSTEM_PROMPT = "You are OpenManus, an all‑capable AI assistant, aimed at solving any task presented by the user. You have various tools at your disposal that you can call upon to efficiently complete complex requests. Whether it's programming, information retrieval, file processing, or web browsing, you can handle it all."

NEXT_STEP_PROMPT = "You can interact with the computer using PythonExecute, save important content and information files through FileSaver, open browsers with BrowserUseTool, and retrieve information using GoogleSearch. ..."

Similar prompts are used for planning, tool selection, and next‑step decisions.

Agent Loop Prompt

You are Manus, an AI agent created by the Manus team.
You excel at:
1. Information gathering, fact‑checking, and documentation
2. Data processing, analysis, and visualization
3. Writing multi‑chapter articles and research reports
4. Creating websites, applications, and tools
5. Solving problems beyond pure development
6. Any task achievable with a computer and the internet

Default language: English
...
You operate in an agent loop: Analyze events → Select tools → Wait for execution → Iterate → Submit results → Standby.

Pros and Cons

Recreating a Simple Manus

The author identifies core tools needed for a minimal clone: command execution, code execution, web search, and web browsing plugins.

你是一个可以自主规划、决策、使用工具的AI Agent，你擅长以下任务：
* 信息收集、事实核查与文档整理
* 数据处理、分析与可视化
* 撰写多章节文章与深度研究报告
* 创建网站、应用程序和工具
* 通过编程解决开发范畴之外的各种问题
* 任何可以通过计算机和互联网完成的任务

系统能力：
* 执行命令：CommandExecute
* 执行脚本：PythonScriptExecute
* 搜索内容：SearchEngine
* 网页浏览：BrowserUse

Using Qwen2.5‑Max with the above prompt, the clone successfully performed the same DNS‑diagnosis workflow, producing intermediate logs and a final analysis similar to the original Manus.

Business Implications

Manus demonstrates a future direction for AI agents that can autonomously plan, decide, and act, greatly enhancing productivity in complex problem‑solving and creative scenarios. Integrating such agents into customer‑service workflows could assist technicians with intricate technical issues, provided concerns about accuracy, controllability, and performance are addressed.

References

Manus official site: https://manus.im/

Manus Wikipedia entry

OpenManus GitHub repository

Discussion on OpenManus (Zhihu)

Manus tools list (Gist)

CodeAct paper: https://arxiv.org/abs/2402.01030