How Manus Redefines the Universal Agent: An In-Depth Look at Its Workings and Interaction Designs

Recently, one of the world's first general-purpose intelligences (Agent) Manus Manus has released a preview version, and the official results are shocking. Unlike many AIs that only stay at the "suggestion" level, Manus not only shows strong task planning capabilities, but also realizes a qualitative leap in task execution, truly achieving a closed loop from planning to execution. So how does Manus work? In this article, we will bring you an in-depth understanding of Manus' Workflow, Memory, and Frontend Interaction, and analyze how it integrates computer operations, in-depth research, and coding agents to achieve the goal of "Less is more" intelligent emergence.

I. Say goodbye to paper: Manus's "plan-execute-update-deliver" workflow

While many AI assistants excel at planning but struggle to put it into practice, Manus takes a different approach, seamlessly moving from planning to execution in a way that more closely resembles human work habits. At its core, Manus creates a Markdown-formatted list of tasks (todo.md) and manages the entire task lifecycle through this list. This approach is much more intuitive and efficient than many systems that manage tasks through the context of a planning agent.

As shown in the figure above, this is an example of a todo.md file for planning a "7-day Japan trip and proposal plan". It not only lists the tasks to be completed, but also marks the completion status of the tasks with "[ ]" and "[x]". This is not only intuitive and clear, but also easier for the Agent to manage and update, making it the "memory" of Manus.

1. Planning: it all starts with todo.md

Manus' workflow begins with an exhaustive to-do list. This list, in the form of a Markdown file, is not only the starting point for the task, but also the vehicle for the Agent's memory. The user needs to list all the tasks in as much detail as possible to provide Manus with a clear guide to what to do.

2. Implementation: computerized operations, in-depth research, coding agents, a three-pronged approach

With a clear list of tasks, Manus began to tackle them one by one. In doing so, Manus demonstrated a strong combination of computer manipulation, in-depth research, and coding agents.

• in-depth study: Manus has powerful information retrieval and web page interaction capabilities. It can search a large number of web pages at once (23 in the demo) and simulate various user actions in the browser, such as scrolling, clicking, and so on. Each step is recorded in a screenshot, making it easy for users to retrace their steps.
  • Browse:
    
  • Scroll down:
    
  • Click:
    
• computer operation: Manus is able to interact with the operating system of a virtual machine, execute terminal commands, manage files (creation, deletion, modification), operate a browser, and realize real "computer use".

  Manus executes terminal commands

  

  Manus Managing Project Documents

coding agent: For coding tasks, Manus gives them to specialized coding agents. The effect is said to be similar to using the Claude models, capable of generating high-quality code (e.g., HTML, Python, etc.).

HTML code generated by Manus

3. Update: real-time tracking, progress at a glance

As tasks are executed, Manus updates the todo.md file in real time, marking completed tasks with "[x]". This way, the progress of the tasks is clearly recorded, and the user has a clear picture of the status of Manus' work.

Manus updates todo.md file

4. Delivery: results at your fingertips

Manus generates final deliverables when all tasks in the todo.md file are marked as complete. To enhance the user experience, Manus also provides a specialized session file management interface for users to view and manage the generated files.

Deliverables generated by Manus

Manus Session File Management

More than "remembering": Manus's self-learning memory mechanism

Manus not only remembers user commands, it learns from them. Its unique knowledge and memory mechanisms allow it to learn user preferences and best practices for specific tasks and automatically apply those lessons when similar tasks are encountered.

This means that users can continually improve their productivity and accuracy by "teaching" Manus how to handle specific tasks. For example, you can instruct Manus to summarize the results in a table when working on a resume, and Manus will do it automatically the next time it encounters a similar task, without having to repeat the instruction. This ability to "learn by doing" is what makes Manus so smart.

More than just "works": Manus' ultimate interactive experience

Manus is not only powerful, but also has a great user experience. The smooth output effect of session playback and the real-time progress tracking on the right side let users know the working status of Manus at any time, as if they have a "visible" AI assistant. This design not only enhances the user experience, but also strengthens the user's trust in Manus.

Manus session interface with real-time progress tracking

IV. Summing up: less is more, intelligence emerges

The Manus team upholds the philosophy of "less structure more intelligence", which means that through quality data, powerful models, flexible architecture, and solid engineering, computer operations, deep research, coding agents, and other capabilities emerge naturally instead of simply piling up features.

Manus combines computer operations, in-depth research, coding agents, and other technologies to realize a truly closed loop from task planning to execution through simple and efficient Markdown task management and excellent front-end interaction design. This "less is more" design philosophy and the breakthroughs in the field of general-purpose agents may be the reason why Manus dares to claim "redefining general-purpose agents".

This article is mainly based on the official demo to analyze, there may be understanding of the deviation, readers are welcome to exchange corrections, and jointly explore the future development of the General Agent.

About Manus is what, many big brother last night also had a discussion, in fact, the problem is very simple, Manus release the first practice has been carried out to explain the principle:Manus What exactly is a Universal Intelligence?and AIGCLINK's answer is essentially the same.

AIGCLINK's view on Agent