My thoughts after testing Manus

After thinking about it, I decided to write an article about my own experience.

Not a simple list of features, not superficial acclaim,

Don’t pay attention to the polarized evaluations.

It's about sharing some real feelings and thoughts.

Let's see how Manus performs.

In one sentence, Manus has done a good job in engineering. The upgrade from "conversational suggestions" to "autonomous and controllable closed-loop execution" allows users to intuitively feel the autonomy of the AI ​​Agent and how the agent's agency is concretely manifested. It is no longer just answering, but taking action. It is no longer just suggesting, but executing.  It has pushed the Agent form forward a small step. Of course, the disadvantage is that the context length is still not enough to complete complex tasks, and some capabilities need to be optimized.

The evolution of agents: from conversation to action

I don’t know when we started to get used to talking to AI and letting big models serve as knowledge bases and thinking partners.

They are getting better and better at understanding our questions and giving insightful answers.

However, the essence of conversational responses is consultation, not execution .

The big model gives you suggestions, but the execution still needs to be done by you.

In the past, we solved the problem by building intelligent agents and manually choreographing workflows.

Manually arrange through prompt+workflow+plug-in+various API calls,

This method is very difficult for more ordinary people.

The emergence of Manus allowed me to experience for the first time the upgrade of Agent from "dialogue suggestions" to "closed-loop execution" .

There have been many articles explaining what Manus is, such as Yize's Manus dispels the fog between people and agents | 8 hours of live test. I will not go into details about my real experience with Manus here. We can just think of it as AI manipulating a virtual machine & browser without a graphical interface deployed in the cloud, sensing the computer environment, and performing various operations.

Give it a task, and it will independently plan, find resources, carry out actual operations, and ultimately deliver a complete result.

It's like we have a multi-talented intern who never sleeps.

You no longer need to worry about the middle steps, just enjoy the final results.

This is a new way of interaction and a new working relationship .

This change is not just a pile of functions, but a fundamental leap in the interaction mode:

• Before, AI was your conversation & thinking partner
• Now, AI is your executive assistant

Without further ado, let me get straight to my actual testing and let the results speak for themselves.

My actual test case

In order to cherish the opportunity, I used the "standard" mode (rather than the "high investment" mode to explore the boundaries of Manus's capabilities) to conduct task tests in multiple different fields.

I wonder how far it can go in real-world scenarios.

1. Analysis and Research

1. Attention is becoming increasingly scarce. How does FOMO affect our attention allocation?

This task involves academic research in a professional field, and I want to test Manus's ability to handle academic problems.

• Test link : https://manus.im/share/HqAtcKFZi6N1asxd8libIZ?replay=1

• Test prompt words : In order not to affect the reading experience, you can directly look at the picture

• Test results : The task completion rate is relatively high, and according to the evaluation, it is a work with a score of 80.

It first browsed the relevant information, collected materials, sorted and summarized them, modified the writing logic, and finally edited the text output. The feedback from the group member @韩跑跑 who proposed this case:

• The format and chapter distribution of the research report is basically a thesis. If it is slightly modified and some of your own research data and results are added, it can become a rough undergraduate thesis.
• The references given in it are real and the corresponding papers can be found on the Internet (this is better than OpenAI's DeepResearch, because many academic references using DR are not available)
• The research and analysis of fomo is very impressive, but the subsequent suggestions for response strategies are a bit general and not in-depth.

One-sentence evaluation: This is really a very good tool for academic research or learning of some scientific knowledge (for example: the neural mechanism of fomo; the related physiological manifestations of depression, etc.). It greatly saves your own search time and is an efficiency assistant.

2. An empirical analysis of the impact of FDI on industrial upgrading in the Yangtze River Delta region

The second test case is economic analysis of a specific region, a task that usually requires specialized data collection and analysis capabilities.

• Test link : https://manus.im/share/SYfLdFC6b46xEyyrs2LquK?replay=1

• Test prompt words :

  "I want to write an article titled 'A Study on the Impact of FDI on Industrial Upgrading in the Yangtze River Delta Region' to study the impact of foreign investment on industrial upgrading. Please combine all online information, analyze and output the research report, and complete the entire process of empirical analysis of the paper."

• Test results : Mission accomplished.

Manus first collected the information, then conducted data analysis, designed the research methods, completed the empirical research, and finally wrote and improved the entire paper.

Of course, we can see that there are actually two tasks in its task list that are not completed. I am not sure about the specific reason-.-.

The specific completion degree and score of this information needs to be scored by everyone.

2. Daily work

1. Monica product research and analysis

One of the most frustrating things for Internet professionals is that they often need to research and analyze new products. This time I wanted to test Manus's ability in product analysis.

• Background : Some positions often encounter a new product and need to quickly research and use it, or research new product features and updates. This type of work can be completely handed over to Manus.

• Test link : https://manus.im/share/9ILnKYG7XX0AtRCDjHCa70?replay=1

• Test prompt words :

  "I have a new AI product called Monica. Please help me write a product analysis report. Pay attention to finding its strongest capabilities and matching it with which tasks it can solve. Compare it with previous solutions, what are its capabilities, and what other similar competing products are there. Please form a complete report."

• Test results : The quality of task completion exceeded expectations.

Manus independently conducts information collection -> competitive product research -> comparative analysis -> report writing -> final delivery .

I looked at the data and information related to competitors and found that they were relatively accurate. I asked my friends and they said that it can basically meet their basic demands, with a score of 70.

2. Convert the dialogue transcript into a short video

Another common requirement is to distill long videos or long text content into concise short video scripts.

• Background : Several hours of interview transcripts need to be refined into a 15-minute video script, which involves issues such as information extraction and style unification.

Private data will not be shared with others for the time being.

• Test results : The task completion rate is relatively high.

Manus first extracts and analyzes the content, then creates the video script (including rough cuts and fine cuts), and finally integrates all the content into a complete video script.

It even gave some golden quotes from the interview record (80 points), but in terms of the completeness of the final theme, it felt that it only read the first half of the content and lost some of the second half (50-60 points).

I think this has a lot to do with the length of the context, and it was written all at once. I didn’t have the time or the bother to argue with him, which would have the opposite effect.

3. Course polishing

1. Course Design on the Theme of "Management - Task Allocation and Authorization"

A practical question from a friend, I would like to test Manus's ability in the field of education, especially in curriculum design.

• Test link : https://manus.im/share/2Z9Jv3tm8rpEn2WXHkYLpi?replay=1
• Test prompt words : In order not to affect the reading experience, you can directly look at the picture

• Test result : The idea is correct, but the final result is a failure.

The reason for the failure is that the context is too long for the system to process.  This is also the biggest problem of Manus at present. The limitations of the basic model make it impossible to complete more tasks, and insufficient resources make many scenarios impossible to realize.

From the picture we can see that the first half of the content has been completed. I showed it to a friend and he said it was beyond his expectation. The quality was over 75 points.

2. Design of DeepSeek related internal training course system

This is another course design-related test, the purpose of which is to collaborate with Manus to polish a high-quality course system.

• Test prompt words (overview):

  "I am currently working on a DeepSeek AI internal training course. The course content is as shown in the attachment, background XXX, goal XXX. Now I feel that the course is not systematic enough, and the logic between the upper and lower parts is rather confusing. Please help me optimize the entire course outline based on the above background information. The content must be rich and logical enough!"

Due to privacy concerns, I won’t post the link yet.

• Testing experience : I executed it twice. The first time, I asked it to help me output the PPT, but the context was too long and I had to stop . The second time, it conducted research and analysis on its own and provided an optimized outline. I kept interrupting it during the process to make adjustments, and finally got a course outline with a score of about 80 points, which greatly reduced the time cost of adjustment.

4. Fantasy

1. How to earn 1 million?

I also wanted to test some abstract problems to see how Manus handles these kinds of open-ended tasks.

• Test link : https://manus.im/share/2Z9Jv3tm8rpEn2WXHkYLpi?replay=1

• Test prompt words :

  "How can an ordinary person earn his first 1 million yuan in a short period of time with 100 yuan? Please give a practical and correct method, exhaust your thoughts, and collect all relevant information on the Internet."

• Test results : The task was completed, but it was interrupted once due to resource problems . After retrying, the result file was finally output, and a feasible path was given in stages and periods. You can try to see if you can earn 1 million. . .

2. Manus plays Red Alert?

After seeing Yize testing the cases of 2048 and Pokémon, I really want to let it play classic games like CS or Red Alert in a headless environment, but the number of times is not enough...

Let's test it another day and see how it performs..

Some thoughts on Manus

1. Abandon unrealistic fantasies

During the actual testing process, my most profound experience is:

Once you have such an agent, you can ask it to do something.

How to ask a good question, how to realize an interesting scenario, how to solve a pain point problem, is what everyone needs to consider...

I see many people put forward vague demands such as "make a fun game" and "make an awesome website".

It is just like making a wish to have a child. Whether it is a "magic pill" or a "magic pill" that comes out depends on luck.

This type of testing is a misuse of the tool.

The truly effective way to use it is to propose clear, specific, and bounded tasks.

Even the most powerful tools require clear guidance to achieve their greatest value.

2. Some issues to be resolved at this stage

• The system breaks down when the context is too long, and still faces bottlenecks when processing particularly large datasets or complex tasks.
• In some test cases, the task was interrupted due to resource problems and needed to be retried to complete.
• For tasks that require multiple levels of deep logical reasoning, especially those that require establishing complex connections between multiple fields, Manus sometimes takes a relatively conservative approach and is not innovative enough.
• When a task is interrupted, the recovery mechanism is not perfect. Ideally, it should be possible to continue from the breakpoint rather than restart, especially for complex tasks that take a long time.

Wait, we need to correctly understand some of the current limitations of Manus in order to better apply it. There are some problems that both large models and Agent classes will face. We look forward to having good solutions in the future.

3. The greatest value of Manus

Many people said that the technologies used by Manus had no barriers and could even be replicated in a few hours, which caused all kinds of public opinion.

But I think,

If we say that the release of DeepSeek has allowed more people to understand AI and its capabilities.

I think the biggest contribution of the release of Manus is that it can allow more people to have a clear understanding of the concepts of Agent and intelligent body, and once again broaden the cognitive boundaries of AI capabilities known to the public.

For most people, you can use Manus as a real intern.

But at the same time, we should not overly "deify" or "stigmatize" it.

The release of Manus will surely give birth to a new generation of AI Agent application ecology .

Major manufacturers will also speed up their pace to catch up with this trend.

This may also be the dilemma that Manus may face. Universal Agent is almost the only way for large companies to move forward.

4. Less Structure, More Intelligence

Manus minimizes manual control of the model and only needs to lay the groundwork.

Let AI play its own role and internalize relevant capabilities.

This is also a non-consensus that is discussed a lot in the industry.

For example, Flood Sung expressed his opinion when Kimi released k1.5.

“The current Agentic Workflows are just things with Structure. They will definitely limit the model capabilities and have no long-term value. Sooner or later, they will be replaced by the model’s own capabilities.”

Manus is designed in this way. There is no built-in workflow. All capabilities are naturally evolved from the model rather than taught by workflow.

The decision-making ability of a startup team is that it is easy to reach consensus, which is its biggest advantage over large companies.

5. Work and life in the Agent era

With the development of agent technologies such as Manus, the way we work will definitely change:

1. Redefinition of responsibilities : Repetitive and mechanical information processing tasks will be largely handed over to AI agents, while humans will focus on creativity, decision-making, and interpersonal interaction.

2. Skills for managing AI : One of the core competencies in the future will be how to effectively manage and guide AI Agents, just like managing a team.

3. Parallel processing capability : By deploying multiple agents to handle different tasks at the same time, work efficiency will be greatly improved.

As the new metric "Agentic Hours per User (AHPU)" proposed by the Manus team reveals,

Future productivity will depend on how long users can effectively delegate tasks to AI.

Conclusion: Less malice, more beauty

The biggest difference between humans and animals is that humans can use tools.

Now, we have given AI the ability to use tools.

Now, the emergence of Manus has put us at a new crossroads——

When AI is no longer just a static tool but an assistant that can "think" and "act", where will our relationship with technology go?

Like a child learning to walk, our digital companions are learning how to move autonomously.

They’re not perfect and they occasionally fail, but each attempt redefines the boundaries of human-machine collaboration.

The significance of Agent may not be what it can do for us, but what it allows us to focus on?

Perhaps, the real value lies not in the technology itself, but in the space it creates for humans——

To think, to create, to feel, to connect with others on a deeper level.

Technology is not just a means to achieve an end, it is also a way of "revealing", revealing the world while also revealing ourselves.

While creating an AI that can "do things", the Manus team took the agent form a small step forward.

I hope to be more tolerant of start-up teams.

I hope there is less malice in the world.

More beautiful

Note: All test links and experiences in this article are based on the Manus standard mode, not the high-investment mode. Some test links are not disclosed for privacy reasons.