To be honest, I still don't have an invitation code for MANUS, so if you insist that you can't have a say without experiencing it, you can just close this and save yourself ten minutes.
I want to discuss two main points - one is about the marketing controversy surrounding MANUS, and the other is about the clever tactics used in MANUS' product.
Product: No Breakthrough, But a Breakthrough
MANUS has not achieved a breakthrough in technology, and this may be the biggest consensus after the controversy. The core validation case comes from the MetaGPT team (who work on the programming agent team), who replicated an OpenMANUS in three hours.
But without a doubt, the product MANUS has brought to everyone is shocking. It uses AI as the "human hand" to automate a series of processes, including a large amount of autonomous information collection and browser interaction. These will ultimately be packaged, and users only need to tell MANUS what they want to do, and they can just pull up a chair and watch the show.
If you, like me, follow the dynamics of the AI field on a daily basis, you will find that MANUS, compared to the innovations in model training of DEEPSEEK, is more like a Frankenstein's monster:
1. Task sorting and knowledge base calling. Most models have this capability, and from the perspective of prompt engineering, pre-sorting the tasks will improve the final output of the AI.
2. Information collection, sorting, and analysis. Also known as Deep Research, this is now also gradually supported.
3. External tool function calls. Whether it's MCP or a large number of open-source tools like browser-use, they have already been incorporated into daily use scenarios.
4. Multi-agent collaboration. It's been about one or two years since Devin (I don't quite remember the exact time). The MetaGPT team that replicated the product in three hours is doing multi-agent collaboration in the programming domain.
Many people cling to the fact that MANUS has no innovation and start to mock and ridicule any praise.
You must know that arrogance is the greatest enemy of progress.
Why don't you ask yourself: since they are all existing things, why did the integration suddenly explode? Since they are all existing things, why don't you sew them together?
The perspective of the public on innovation is still too narrow, always clinging to technology and not realizing that innovation in product concepts and business models is what has a more far-reaching impact.
I've included an article at the end that does a great job of understanding the MANUS product (because it's a recounting of the team's own words):
This is also the main reason I wrote this article. I feel that the points mentioned here are worth pondering for anyone who does product work.
Let me talk about the points I see from this that are worth learning from MANUS:
1. Interference Resistance of the Intelligent Agent Workflow
Similar concepts are mentioned in game design, such as "flow".
That is, the user's experience of using a product is composed of a series of behavioral processes. For example, when I use Alipay to pay my phone bill, I need to open Alipay, click "Recharge" on the home page, select the phone number, select the amount, confirm, pay, and receive the recharge success feedback. If there is an external factor interference in this process, it will disrupt the overall process and cause a decline in product experience.
The existing AI agents, although not as severe as the above example, where the user's workflow is completely incompatible, do have the problem of occupying the work space. For example, in the case of AI reading web pages, the AI provides an understanding based on the current web page, but you cannot modify the current web page, otherwise it will become invalid. Otherwise, you have to manually copy the link and give it to the AI in a separate window. Another product I recently tried, same.dev, which focuses on directly copying the front-end source code, has the problem of interrupting the operation by bypassing the web page, literally occupying a work space, and the experience is very poor.
Of course, it's not about saying which one is better. But the changes in the MANUS product have significance for us to think about the evolution of the AI agent form.
From embedding AI in the browser to embedding the browser in the AI interaction page, the former satisfies the user's daily usage needs (AI is just a part of the work and life process), and the latter satisfies the needs of the intelligent agent (reducing interference from irrelevant factors).
For AI Agent products, what is exposed to users should only be an interaction page, receiving input, displaying the process, and outputting the result. The process can only be displayed, and the user will not be interfered by the process.
I'm also reminded of the problem I mentioned earlier with the OKX wallet embedding a browser. You can't let the user actively interrupt the process to achieve the goal.
Disruption of "flow" is a very poor experience.
2. Rethinking "External Tools"
In the eyes of most people, only standard protocols like MCP or pre-packaged libraries may be considered "external tools".
In fact, so-called tools are those with stable input and output, allowing users to have a clear expectation of the black box.
Programming, as the application scenario of AI with the highest certainty, has countless scripts and countless modules that can be packaged as tools.
It's fine to compete in model training to solve math problems, but if you try to directly tackle math problems from the model application perspective, that's just stupid.
Why not write code to solve it, why bother with some vector mapping nonsense?
MANUS' design of the general-purpose Agent actually points out a key point: don't try to use AI to directly solve all problems.
AI is just a hand.
The hand uses tools to solve problems. Tools can be pre-defined or temporarily written.
Human - Hand - Tool - Task.
Who cares how many layers of tools are in between?
3. Non-destructiveness to the Use Scenario
Programmers must have a deep understanding, in different projects, the dependency library versions are a mess, as long as one of the key library versions is not matched, the project may run with errors.
This is also the necessity of virtual environments in programming. npm based on project dependencies installation, python creating virtual environments, docker containers, etc., in my understanding, are all to ensure the independent customizability of the environment.
This may be the consensus of all product-level intelligent agents targeting the consumer market at this stage: don't occupy the user's local environment, use the cloud.
There are products like bolt.new and mgx.dev that have chosen to run and write/debug directly in the cloud. But these are all programming-type agents, lacking a general-purpose agent counterpart.
In stark contrast to MANUS' approach is another product, Highlight. After you download and install it, a floating window will appear on the desktop, with some AI-combined operations for applications based on the current work space.
Looks eye-catching at first glance, right?
For example, if I can't do web scraping, I switch to a browser page and let Highlight do the web scraping for me?
From my personal experience, Highlight's approach has already disrupted the original workflow, because I have to keep the process running to let Highlight operate. But in reality, anyone in their work will switch between pages back and forth, and can't just wait for the AI to finish. Also, if the AI uses my browser to scrape the web, will it be using my IP, and will it affect my future access?
The local scenario has been destroyed.
These points may seem easy to think of just by talking about them. But how to systematically carry out a series of designs, I think it is still worth experiencing.
Finally, let me say a word about the upper limit of MANUS.
My expectation is to control expectations - AI cannot do everything for us. Even if future smart homes introduce AI modules and become "tools" that the "human hand" can use; even if more and more "tools" appear on the desktop that allow us to control production-level software with just natural language, there will still be many times when humans need to do process control and verification, because AI's cognition of the world is built on a huge black box, and they will have "illusions".
In any case, since we believe that human truth is greater than AI truth, humans need to verify the production of AI. And the more we hand over to AI, the more content needs to be verified. This will ultimately balance out to a boundary.
This boundary is the upper limit of the general-purpose Agent product.
Marketing: Not Afraid of Controversy, Afraid of No Controversy
Looking at the MANUS outbreak, rather than saying it's a spontaneous controversy within the circle, it's better to say that there are two "hands of God" guiding it behind the scenes.
These two hands of God are likely to come from the official side as well. Guiding by hand, isn't that wonderful?
First, I had Grok summarize the major marketing events of MANUS during this period:
It can be clearly seen that the official core slogan is "the world's first universal AI Agent".
This is a very controversial statement.
A. For outsiders, this statement is very eye-catching;
B. For me, a daily follower but not an insider, it is clear that this is a word trick: as I have introduced above, the product is a Frankenstein, and in no way can it be called the first; but if the word "daily" is added, there has indeed not been such a consumer-level product claiming to handle general tasks that has caused a lot of public opinion. And since the slogan is originally inclined to exaggeration, it is also understandable to say so;
C. For insiders, I think most of them may be quite indignant, after all, the results they have been working hard on have been sewn in, or in their eyes, this thing has little technical content, but it has stolen the limelight.
——There is a conflict of positions.
Where there is a conflict of positions, there is debate, and the public opinion will eventually ferment, bringing a leverage effect to the promotion of manus.
Don't look at the surface, look at the results.
The result is that manus has gained global attention. The cost-effectiveness of this marketing is maxed out.
And the invitation code mechanism.
While attention is fully engaged, the invitation code is strictly limited, on the one hand based on cost considerations, and on the other hand to cover up product deficiencies. After all, the "universal" claim has been made, and once opened, it will immediately be overwhelmed by various bugs and feedback from the A-class group. That would not be a conflict of positions, but a one-sided situation, and it would likely cool down. From this perspective, the invitation code mechanism is actually a kind of early testing, with as many dev bugs fixed as there are codes released, allowing seed users to help improve the product first.
There is also the plan of hunger marketing. Hunger marketing is essentially about competing for attention, and anyone who uses invitation codes will inevitably be cursed by the "love but not get" people, which is very normal.
What happened afterwards, the X account freezing incident, the technical jailbreak, and the open source, were seen by many as the evil consequences of over-marketing.
On this point, my personal view is: the AI circle is still too stiff. I suggest bringing in the crypto circle and learning some shameless spirit.
The manus team is actually not bad, after all, the previous product Monica has been stably profitable, little knowing that many small teams are still struggling on the edge of subsistence. At this time, if you tell me not to over-market? If you can achieve high marketing effectiveness at low cost, I'd like to ask why not?
Can face support the R&D team to continue? Can face provide enough funds for innovation?
Face is worthless. In this era of entertainment to death and information explosion, attention is worth money.
The crypto circle is the most cruel and also the most tempering place, because it is too close to money itself, so what you can see is the most real human nature and the most bloody tricks (not the tricks are bloody, but you are bloody).
One must have principles, but you cannot fail to understand and accept some realities that are not your principles.
Otherwise, you will die a miserable death.
References
Full Replay: How was Manus born? | Geek Park
Behind the Manus Explosion, How Can Agentic AI Products Build Lasting Competitive Advantage?
