GuShiio.com Wen| hengheng
Manus has been surfing the screen all night, and my circle of friends is crazy about this Agent that can control the computer independently. It can not only help users answer questions, but can even break down tasks based on the user’s questions, make plans to solve problems, and automatically execute them.
Manus is not a technological breakthrough in the pedestal model, but a creation in the actual implementation of AI, increasing application scenarios and suitable groups of people.
It is generally believed in the industry that answering chat robots can at most provide some simple work assistance and perform complex tasks. In the end, they still need to build a workflow of AI Agents.
However, there are thresholds for the construction of Agents, not only technical thresholds, but also cognitive thinking and structural capabilities. Not all users have the ability to break down work into detailed SOPs. Under Manus’s logic, the most difficult task of creating a workflow SOP is done by AI. Therefore, through simple natural language responses, each user can obtain results close to the output of Agents, which is almost directly available.
The articles on the whole network today are mainly a brief introduction to the Manus function. Friends who are familiar with me know that I have always advocated not only analyzing the product, but also digging deep into the logic behind the product and its corresponding trends. Therefore, in this article, we will conduct in-depth analysis and prediction of AI products and AI implementation based on Manus.
Agent enters its third phase, productivity change is coming soon
The characteristic of Manus is that it can disassemble the workflow independently, and then invoke the most suitable large model at each step according to the order of disassembly to process tasks.
It’s like a very talented J colleague. Every job will be broken down into a detailed execution plan in advance and then strictly implemented according to the plan. The final effect will definitely not be bad.
For example, if Manus organizes an organizational chart of OpenAI, it will not directly rush to do this, but will first perform a task disassembly to clarify what specific and detailed steps are needed to do this well. Then, according to the steps you have disassembled, you will automatically execute the tasks item by item, and finally produce results that are directly available to users.
It is so important to use this directly. If you have really used AI in depth in the past, you will know that the previous chat-style AI seemed to provide a large amount of replies, but if you actually use it, you will find that many places are not in line with reality and you will have to be changed manually.
Therefore, Manus’s working logic is actually an Agent, which solves specific practical problems by completing the disassembly of the workflow.
The Chinese name of an Agent is AI agent or agent. To put it bluntly, it is a robot that can perform practical and professional work according to the workflow, solve practical specific problems, and let it perform various specific work according to the process suitable for the company’s business., just like a real colleague.
If you just have a simple conversation in the AI dialog box, the work that AI can do is very limited, and the output of AI is not very controllable. Returning to our actual work, any work has a process and standard limitations, so that the output results can reach the predetermined standards.
For example, if AI is allowed to work directly in the work of “writing a little red book for the company’s business”, the results will be basically unavailable. Because the proposition is too big, the operation of the little red book includes a variety of different types of work. Link, it is difficult to complete a relatively complex job in one sentence.
If you want AI to do this job well, you actually have to specify the working steps for AI:
- Learn the knowledge base of the company’s business and understand the company’s business situation and past content
- Collect today’s hot news related to your business through a large network-enabled model
- Analyze what topics can be combined with business through a large model
- AI generates copy for topic selection
- Use the large model to accompany the copy to create a Prompt of raw pictures
- Call the large model of Wensheng diagram and use Prompt to generate a matching diagram
- Manually post copy and pictures to Xiaohongshu
- Grab likes and collected comments data through published notes links and conduct data analysis
In each link, AI only solves one specific problem, and then strings together many AI tasks to become an AI workflow.
Therefore, AI for conversation and chatting can only provide some inspiration at best. To truly solve productivity problems, we must rely on AI Agents that integrate SOPs.
This is also my consistent view. Collaboration systems are a tool for human society to change productivity. Only through workflow can AI be perfectly integrated into human production relations.
The emergence of Manus means that the development of Agents has actually entered the third stage:
In the first stage, answering robots like ChatGPT solve problems one by one by one. In this process, humans are still the leader. Whether AI can be used well depends on the level of the questioner. Only a questioner with very clear ideas and strong structural abilities can allow AI to solve practical problems.
The second stage is to build an Agent that executes step by step through an agent platform such as Button. This stage is also the main way to implement the current AI application layer. Since not everyone can solve practical problems through AI Q & A, we will fix the steps that can solve problems, let AI execute the verified steps one by one every time, and integrate AI with multiple abilities, let AI only do things in what it is best at. In this way, as long as one person succeeds, others do not need structured abilities and can achieve the same effect.
The third stage, which saw dawn this morning, is the universal Agent represented by Manus. The process of disassembling the workflow also allows AI to replace it, and users return to simple question and answer mode. But the problem is simple, and the process of AI execution is not simple. Universal Agents can complete the problem with high quality in accordance with the workflow, so that the results can be produced and reached a usable state.
If most people could only join in the popularity of AI models in the past, there is no doubt that universal Agents allow all ordinary people to use easy-to-use AI in specific work to solve real-world practical problems. Only by solving practical problems can AI truly transform productivity and production relations.
AI that distributes AI will become the entrance to AGI
In the introduction of the Manus development team, Manus was called a multi-agent system.
What is a multi-agent system? In fact, multiple large models are integrated into one system, and then a dispatch center calls the most suitable large model according to different tasks.
Several leading big model companies are constantly launching stronger big AI models. One day, a big model just reached the top, and the next day, it was surpassed by the next big model. The city wall changes with the flag of the king, and you sing and I appear.
However, there is still a big gap between pure heap parameters and running scores and actual use. In fact, each big model has its own areas of expertise. For example, Claude is extremely good at coding, DeepSeek’s reasoning model has strong control over Chinese, the big bean bag model is unique in speech recognition, and the flexible video model is very leading.
Since it is difficult to have bucket models that are good in everything in the short term, why not let the large models collaborate based on the actual situation of the task and let them only do what they are good at? This is the source of multi-agent thinking.
In fact, there have been such precedents in previous explorations and attempts.
Not long ago, Berkeley built a small model with only 7B parameters and scored 1400 points in the Arena Global Model Ranking. This score was just that Gork3, which used 200,000 graphics cards to train, has just reached this level.
This small 7B model is actually a classifier. It only does one thing, which is to classify and filter the questions sent to it by users, and then direct “big models” such as GPT, Gemini, and DeepSeek to work. It can be called a leader of the AI class.
This is actually Manus ‘logic. Through a super classification system, large models that are good at different fields are collected and processed together to solve users’ practical problems to the greatest extent possible.
The future of AGI must not be a single model that dominates the world, but a good division of labor and integration. In essence, the organizational structure of a modern company is to build a machine-like collaboration structure. Relying on an AI to do all the work is tantamount to relying on an employee to support the entire company.
In this transformation, China will stand at the center of the world
This is not the brainless hype of marketing accounts, but a judgment based on logical reasoning.
The logic is two:
1. There are infinite accumulation of application scenarios and business models in China. First, AI has real value only if it is applied in the field. Secondly, the initial AI may be born in the laboratory, but only in real practice can we create stronger and stronger AI.
2. Decades of hard sowing have finally borne fruit today. National basic education, across thousands of years of history, has never been at its present height and depth. The foundation of AI is still talent. After decades of basic education, China’s talent density has surpassed that of most countries in the world, and it is completely ready to compete with the United States.
Many things are the result of qualitative changes caused by quantitative changes.
At the end of 2022, ChatGPT was born and shocked the world. At the same time, this also means that the large-scale model route of AI technology completely replaces the AI route of rule algorithms and knowledge maps in the past.
China was in an uproar. The United States on the other side of the ocean was actually so far ahead in the AI arms race. At that time, most people were pessimistic about the development of AI in China. Computing power graphics cards and technology blockade, the future of China’s AI technology was indeed uncertain.
But standing in the early spring of 2025, this haze has been swept away. The development of AI in China has attracted worldwide attention. DeepSeek, which represents the capabilities of the pedestal model, has attracted the world’s attention. Today, Manus, which shines brightly in the application layer, also comes from China.
Xiao Hong, the founder behind Manus, is a 2015 graduate of Huazhong University of Science and Technology. He has worked as a “one-partner assistant” and a “micro-partner assistant” before starting a business.
This team also has a more well-known product, Monica, an AI assistant that became popular across the Internet last year. This is an AI application that started as a browser plug-in. Through very keen insight into needs, users can use GPT 4o, DeepSeek R1, Claude 3.7 and other models to do some very specific work in the browser.
Some people say that this is AI in a shell, but Monica’s philosophy is to use cutting-edge AI technology to find a way to implement the application layer to realize the true implementation of AI. The ultimate shell is also a cow?! Under this concept, this team created Manus again.
We are very happy to see that the emergence of Manus is another big step forward in allowing ordinary people to make good use of AI.
I saw in the afternoon that many people accused Manus of being proactive in publicity and promotion. They believed that Manus’s so-called super abilities were marketing and wanted to become the next DeepSeek.
But in fact, Manus and DeepSeek are two different products themselves, one focusing on better implementation applications, and the other focusing on technological breakthroughs in basic models, which are both very valuable. I remember a sentence I wrote a few years ago:
Science popularization is as important as scientific and technological innovation.
This is just an idea, just a starting point. Keep humble and look forward.
