The black and red Manus still has three more passes to pass

author| Focus One Wang Lu

“AI one day, one year on earth.

In the early morning of March 6, Manus made his appearance, which instantly ignited the entire technology circle. It is known as the world’s first general-purpose AI Agent product and was built by Monica.im, a domestic entrepreneurial team. But this technology carnival quickly fell into a whirlpool of controversy. Only two days after its release, Manus’s reputation has undergone several drastic reversals.

Manus officially defines itself as a general artificial intelligence agent that connects thinking and action. It is characterized by being good at handling various tasks. It can not only think (understand complex instructions and reason), but also deliver results (smoothly connect each part of the task and give the final result to the user).

Give a simple example.

If Manus is asked to screen resumes, users only need to send a compressed package containing multiple resumes directly to Manus, and all the remaining tasks will be completed by Manus. Manus will automatically decompress files, browse to find important information, and give candidate evaluation results. It can also make candidate names and important information into tables for convenient user reference.

During this process, users can not only turn off their computers, but also add other tasks at any time.

With Manus, the daily work of ordinary HR seems redundant. This also made many people exclaim that AI cattle and horses replacing migrant workers are really here.

However, Manus belongs to a black and red constitution.

At the beginning of its release, it was dubbed as exploding, punching ChatGPT, standing on par with DeepSeek, and shocking Silicon Valley, and it has unlimited fame. Soon after, doubts followed. Some people believed that it was over-marketing, that it did not have much technological breakthroughs, and that the gimmicks were far greater than the actual capabilities.

Immediately afterwards, news spread that its internal beta invitation code was fired to a sky-high price of 100,000 yuan; its official account on platform X was frozen, indicating that the reason was a violation of platform rules. Subsequently, Manus co-founder and chief scientist Ji Yichao publicly responded that the account freezing may be related to third party mentions of cryptocurrency fraud, and this area has nothing to do with Manus ‘operations.

This series of incidents has deepened the outside world’s doubts about Manus. Is this high-profile product a real technological innovation or a marketing gimmick? How long will the future be popular?

Manus in dispute is really strong or a shell tool?

Manus exploded for two days, and the outside world’s evaluation of it was extremely polarized.

Some people think that Manus has taken AI Agent to a higher level and will soon take away the jobs of migrant workers, while some people think it is just a shell tool.

Let’s see where it becomes a god.

Compared with AI Agents already appearing on the market (such as Doubao and Wen Xiaoyan, which have many built-in agents), one of Manus’s advantages is that it has autonomy and continuity when processing complex tasks.

For example, the user only needs to give a more detailed prompt, and Manus can interpret, disassemble and give the results on its own. It can also conduct self-examination, eliminating the need for step-by-step dialogue guidance steps.

Since Manus runs asynchronously based on the cloud, closing the page after a user gives an instruction will not affect Manus ‘processing of instructions, and the user will be automatically notified of the result after the task is completed.

Another highlight of Manus is that it claims to be the world’s first general-purpose AI Agent product.

What needs to be noted is the word universal, which means that it can handle a variety of tasks, not limited to any field, such as supporting multiple scenarios such as screening resumes, real estate research, and stock analysis.

However, there are also many doubts about these two highlights.

First of all, since Manus is in the internal beta stage and adopts an invitation code system, most users have not actually used the product. Most of the amazing comments on it so far come from the officially released demonstration videos. In view of the fact that many AI products in the past have sample levels greater than actual strength, some practitioners believe that Manus’s true capabilities need further consideration.

Secondly, regarding the versatility of Manus, some practitioners believe that there is suspicion of being a shell.

Some practitioners have commented that Manus does not have the ability to develop large-scale underlying models by himself, and his model capabilities mainly come from Claude from Anthropic abroad. Moreover, many scenarios and functions have already been implemented in other Agents. For example, Devin, as the world’s first AI programmer and an AI Agent that can independently complete coding tasks (learning new technologies, debugging code, and deploying applications), was opened to the public at the end of last year.

From this perspective, Manus is more like adding a shell to various existing AI applications, summarizing applications such as DeepResearch (launched by OpenAI) and Computer Use (launched by Anthropic).

AI software engineer Qin Xiang told “Focus One” that Manus is based on the call of existing large language models and does not involve underlying technological breakthroughs. For example, when conducting stock analysis, Manus obtains data by calling Yahoo Financial API, and then uses programming language to generate visual charts, relying on existing technologies and tools.

AI blogger Automation Hua, who participated in Manus’s closed-door sharing session, also told “Focus One” that the production team shared the three main things Manus does as an AI employee: equipping computers (giving AI the ability to access browsers and tools), assigning permissions (accessing private APIs and authoritative data sources), and providing training (users can adjust AI behavior in real time through feedback).

It can be seen that Manus essentially joins various tools. It can complete complex tasks such as cross-application. It relies on the Claude model and various connected agent tools.

As for whether shell splicing counts as innovation, the industry has different views. Proponents believe that this integration method can quickly realize multiple functions and provide users with a one-stop solution, which has certain practical value; opponents believe that this model is just a simple combination based on existing technologies.

However, practitioners basically agree that from a technical perspective, Manus is far from being the next DeepSeek.

How to treat Manus correctly?

Compared with technology, the key to Manus’s popularity lies in user experience.

The operating steps are reduced, there is no need to repeatedly dialogue with the AI during the execution process, and the interface can be closed, which is not available in previous AI Agent products.

Automation China believes that the significance of AI Agents is to help ordinary people improve their effectiveness. Manus has redefined the value indicator of AI applications from focusing on DAU (User Retention) to AHPU (Agenda Hours Per User, the time it takes users to complete tasks using AI).

“It is this benchmark (which can be simply understood as an evaluation indicator) that determines the product optimization route and distinguishes Manus from other AI Agent products on the market. rdquo; Xinhua said.

He explained that traditional Internet products are all attention economy, hoping to occupy more users ‘use time and ultimately monetize it through advertising, so a variety of vertical Agents are included in the products.

Manus is positioned as a general-purpose AI Agent. When performing the same amount of tasks, it tries to make the user spend as short as possible. Therefore, it has made a major integration, hoping to use faster operation methods and fewer operation steps to improve user efficiency.

This is related to the team’s past background. The team behind Manus developed Monica, an AI browser plug-in assistant in 2023. They found that the browser as a single-user scenario has human-machine control conflicts, and AI frequently interrupts the pain points of the user experience. This led to the conclusion that AI should have an independent cloud browser.

Automation China believes that it is precisely because of this that they proposed Less structure, more intelligence, advocated reducing structural restrictions on AI and relying on the independent evolution of models rather than manual preset processes, that distinguishes Manus from traditional AI Agents.

This is also what most users are excited about Manus. Qin Xiang compared Manus to a digital agent that can automatically disassemble complex instructions and complete multiple tasks for users. It also executes quickly, surpassing Doubao and Baidu’s AI Agents in terms of task execution and user experience.

Qin Xiang also further analyzed the technical principles behind Manus’s core advantages: multi-agent architecture and strong independent learning capabilities.

First of all, Manus has a multi-agent architecture and task decomposition capabilities. It adopts a multi-agent system architecture, which can decompose complex tasks into multiple sub-tasks and collaborate to complete them through different agents. This architecture allows it to efficiently handle long-tail tasks.

Secondly, it has autonomous learning and memory functions, can remember users ‘preferences, and proactively adjust strategies in subsequent tasks. This autonomous learning ability allows it to continuously optimize task execution logic and gradually shift from passive response to active service.

However, when increasing speed, accuracy is also a key factor that needs to be considered.

Manus officially pointed out that it has a low error rate. In GAIA (a benchmark dataset specially designed for general artificial intelligence assistants, mainly used to evaluate the basic capabilities of AI), its scores exceed OpenAI’s DeepResearch in the face of three levels of task tests: basic (sorting files, data analysis), medium (calling API, writing code), and complex (logical reasoning, multimodal processing), but it is lower than the average human accuracy rate of 92%, which is still not as smart as the human brain.

As mentioned earlier, since the product has not yet been widely rolled out, practitioners are skeptical about its actual effectiveness. Many bloggers found in their evaluations that Manus encountered long-time-consuming problems when handling moderately complex tasks; when visiting websites that required login or processing paywall content, it also encountered jams, and users still needed to take over manually.

Manus’s next step is cost, technology, and market competition

Manus is controversial, but he is highly praised. Many people are worried that after Manus is officially launched, if its strength does not meet expectations, it may consume the positive value DeepSeek has accumulated for domestic AI.

Automatic China told “Focus One” that at Manus’s closed-door sharing meeting, officials also mentioned some of the problems they would face. For example, future product strategies will strive to optimize costs and increase operating speed. At the same time, they are also looking for a commercial pricing model that balances costs, so there will be some time before it is publicly launched.

Among these challenges, cost issues are particularly prominent.

“The operating cost of Manus for a single task is about US$2. Although it is only 1/10 of that of competitor DeepResearch, it is many times more expensive than the single task call answer of bean bag.& rdquo; Automation China said that in his view, the computing power called by Manus and the quality of the generated results are different from those of competing products. However, in the context of the current domestic large model products being generally free, this kind of AI can complete high-quality results at one time. The payment model remains to be explored.

Currently, most ordinary users are less willing to pay for AI products, although it is Devin, a powerful AI software engineer, whose high monthly fee ($500) results in very few people using it.

This is also an important reason why Manus needs an invitation code to use it. Practitioners believe that the reason why officials do not fully liberalize use rights may be because of hunger marketing on the one hand, but more importantly, the balance between commercial value and cost must be considered.

With more users, computing power may not keep up.& ldquo; The underlying level of Manus relies on large model capabilities, and computing power constraints can be solved by adding graphics cards. However, many companies cannot afford the high costs behind them, especially since Monica is still a startup company.” rdquo; Automation China said that Kimi encountered similar problems when it was first launched on the market and did not easily choose to add graphics cards.

Secondly, current AI Agents are not as smart as humans and have relatively limited ability to solve problems. Manus is also facing this problem.

For example, the most common situation is that the user’s question is not clearly understood. The same question is repeatedly asked to the AI Agent several times, but the generated answer still cannot meet the standard that matches human cognition.

Because the way AI Agents think when performing tasks is to reproduce mature task processes or simply split problems through training data, in the final analysis, they still rely on the basic capabilities of the large model. Manus has not made a strong technical breakthrough in this regard. Some users pointed out that Manus’s ability to disassemble complex problems is still at a low level.

And the market competitive pressure that Manus is about to face is also considerable.

Qin Xiang believes that Doubao and Baidu’s AI Agents also have their own market positioning and technical advantages. In the future, they may narrow the gap with Manus through technological upgrades or differentiated competition, and these two companies have stronger financial strength.

Many people take a wait-and-see attitude towards the popularity of Manus. More than one investor told Focus One that although they are very optimistic about the development of AI Agents this year, Manus did not surprise them too much due to the appearance of similar AI programmers Devin, and the core technical issues have not been fundamentally improved.

At this moment, compared to Manus, which has a lower moat DeepSeek, many open source replica projects have appeared, and Manus has not much time left.

* The caption and accompanying drawings in the text are from screenshots from Manus’s official website.