When the controversy over “moving headquarters to Singapore and laying off 120 employees” was rampant, Manus’ products did not stop changing. In half a year, it used Chat mode and playbook template library to turn the “blank input box” into an AI Agent that everyone can use, reducing costs by 5 times, increasing speed by 2 times, and automatically fixing video bugs. The actual test of the article found that the moat of Manus is not in the model, but in the rapid polishing of the user workflow. The problem is that the time window left for it is being rapidly compressed by the giants.
Since the beginning of 2025, the wind direction of the AI industry has quietly changed. When everyone is already accustomed to chatbots, a newer concept – AI Agent that can directly complete tasks – begins to be mentioned frequently.
At this moment, a product called Manus entered the public eye. The core problem it wants to solve is how to make AI not only speak, but also directly get started and turn ideas into results. This idea, combined with its use of the whole emotion at the time of its debut, quickly attracted the attention of the global tech community.
Although OpenAI’s Operator released a research preview version in January this year, judging from the direction of the later story, the emergence of Manus for the first time threw an independent and “general” AI agent product to a wider range of users.
The next few months were fast-paced for the company. As soon as the product was unveiled in March, it announced the completion of a $75 million Series B financing in April, led by Benchmark, a top venture capital firm in Silicon Valley. At the same time, it appears in the ecosystem of major Silicon Valley model companies such as Anthropic and has been displayed in a high-profile manner. His founding team began hosting user meetups around the world.
The meeting even went to Nepal
In June, Manus officially confirmed that it would move its headquarters from Beijing to Singapore, with plans to open offices in California and Tokyo.
Headquartered in Singapore, the hub where East meets West, brings it one step closer to the global market. Of course, in the current complex global environment, doing so is also a realistic choice. The acceleration of the move to Singapore is also believed to be in response to US investment scrutiny in the AI sector.
B-end product manager’s ability model and learning improvement
The first challenge faced by B-end product managers is how to correctly analyze and diagnose business problems. This is also the most difficult part, product design knowledge is basically not helpful for this part of the work, if you want to do a good job in business analysis and diagnosis, you must have a solid …
View details >
Accompanied by the migration of the headquarters is the rapid restructuring of the team.
Recently, Manus has adjusted its domestic team, with about 40 core technical personnel being transferred to its new headquarters in Singapore, while the rest of the original domestic team of about 120 people is facing layoffs. Some commentators pointed out that for a startup, the team of 120 people seems a bit “bloated”, and adjustment may be inevitable.
In this way, in just a few months, Manus completed the transformation from a local star startup to a global headquarters in Singapore, which was described by many as “running away”.
01 Under the hustle and bustle, how does Manus make products?
A company’s structure and strategy have undergone drastic changes, and the eyes of the outside world are often focused on its business actions. But for users, the question that is more concerned is always what happened to the product that initially whetted people’s appetites.
To understand Manus, we can use an analogy. In the past, we used large models to recruit a knowledgeable doctoral intern, but we only gave him paper and pen, and what we could do was very limited. What Manus wanted to do was to directly give this “intern” a fully equipped computer, so that he could access the Internet, use tools, and write code by himself, so that he could really do his job. Therefore, its core is intended to bridge the gap between “ideation” and “execution”.
To achieve this, it is designed internally to be more like a micro-project team, automatically planning, executing, and validating tasks. The key to this design is that it allows users to “see” the whole process of AI work. Instead of black boxes and final answers, users see a clear list of tasks that can observe in real time which step the AI is taking.
Although the vision of general AI agents is ambitious, after the initial heat, real problems have gradually emerged. Many users have reported that they are faced with a blank input box and do not know how to assign a specific complex task. At the same time, the operation speed of the agent is slow and the cost is not low, which affects the frequency of user use. Data shows that the number of visits to the product has declined continuously in the early stage.
These real-world problems prompted the Manus product team to make some specific adjustments. For example, in June 2025, Manus launched a free chat mode (Chat) and a template library called Playbook. Chat mode is used to handle simple everyday tasks, while Playbook provides ready-made templates for some common scenarios. At the same time, in order to solve the cost and speed problems, Manus has also optimized the architecture, which is said to have reduced costs by 5 times and increased speed by 2 times in three months.
The core concept of Manus’ product philosophy is called “less structure, more intelligence”. The team believes that too rigid product structures should not be designed for the limitations of current large models. There is a circulating claim that the Manus team itself did not write a preset “workflow”. They believe that by giving the model fewer restrictions, it can learn to find the optimal solution on its own.
This idea determines that Manus chooses to become a “shell” application, which focuses on polishing the application layer, integrating and calling the industry’s best basic large models (such as Anthropic’s Claude series) to complete tasks, rather than building large models by itself. This is a pragmatic choice to deliver value to users faster. But challenges also follow, as a “shell”, its core capabilities are greatly limited by the underlying large model it calls. Once the capabilities of the underlying model are improved, or large companies like OpenAI and Google launch more powerful native agents, Manus may face the risk of being “swallowed”.
At present, Manus is trying to build this barrier through rapid product iteration and deep integration into user workflows. The future plans mentioned in their blogs,Including scheduled tasks、Application integration (email、calendar、Cloud disk), etc., All point to a more ambitious goal, Becoming an indispensable “third hand” in users’ work and life. This vision also echoes its name – Manus means “hand” in Latin.
02 It can still be used like this now
From the initial invitation code hunger marketing mechanism, to the later limited and free opening, to the complete paid hierarchy to officially provide services to all users, what changes can Manus do?
On the contrary, this is not discussed by many people.
Many people’s impression of Manus may still be stuck in the stunning demonstrations of its inception, which allowed it to plan its own travel itineraries or generate market research reports. These capabilities demonstrate its potential as an AI Agent. So, after these months of iteration, what interesting uses does Manus have now? We actually tried it by hand.
The first discovery is that it is trying to find a way to make it easy for ordinary people to use. We opened its “Playbook” template library and chose a “AI Fitness Trainer” template.
This template makes key information such as fitness level, goals, and physical condition into multiple-choice questions, and we set up a user portrait with “knee problems” and need “comprehensive fitness”, click confirm.
After a few minutes, it delivers a web page that can be accessed directly. The content is quite complete, not only with a customized 7-day training plan, but also with special recommendations for knee-friendly low-impact exercises such as swimming and wall squats, and even includes nutrition advice and progress tracking modules.
https://xtboylxo.manus.space/
A similar idea of “templated” is also reflected in PPT generation. We tried to make it “a PPT about a brief history of artificial intelligence, aimed at primary school students”.
It likewise offers a variety of design themes and page lengths for us to choose from.
The final generated PPT has a total of 14 pages, and a few pages have been selected here to post here.
Compared with the PPT function provided by some model manufacturers or ChatGPT tools, it is more likely to provide complete results and reduce the part that users can add by themselves.
Then, we tested its autonomous operation ability and gave it an instruction: “Open Xiaohongshu and comment on the latest note released by Silicon Star Pro: Today Friday, the monkey goes to dance.”
Manus successfully opened the website, but during the login and security verification process, the user was asked several times to “take over” the browser to assist in the scanning and puzzle verification.
I have to say that the latency of this cloud browser is still a bit high, and it took two verifications to pass
In the end, the comment was successfully published, and the IP territory was shown as the United States. The entire process shows that it still requires user involvement and collaboration when it comes to account security and complex verification.
Finally, an important capability update in the model field since its release is in the video field. Manus has also announced an update to its video generation function, and we decided to give it a video creation task.
We asked it to create a dark dish with century eggs and bananas, and made a video.
After receiving the instructions, Manus immediately planned the task steps, created the recipe, generated a reference image, and then called the Veo3 model to generate a standalone video clip with acoustic audio.
Finally, it executed the ffmpeg command-line tool directly in the background, stitching these video clips into a full 30-second short film.
After delivery, we found that the video had no sound. We pointed out the problem, and Manus reacted quickly. It accurately diagnosed that the problem was with the ffmpeg command it was using, and a parameter setting error caused the audio stream to be lost.
Subsequently, it automatically corrected the command, re-executed it, and successfully delivered the final version with sound.
This process of “fixing bugs” is the brightest part of the entire test, which not only demonstrates intent understanding, task planning, and tool call capabilities, but also reflects a certain degree of reflection and error correction capabilities.
03 Write at the end
Looking back at the development of Manus in the past few months, whether it is various controversial actions at the company level or repeated trade-offs and adjustments between the grand vision and user reality at the product level. A central question is still often asked, what exactly is the moat of Manus?
This is a topic that has been repeatedly discussed in the industry. A common view is that as a “shell” application, its core capabilities are subject to the underlying large model it calls and there is a risk of being “swallowed”. This is really its biggest challenge. But on the other side of the coin, there is also a view that the real barrier is not only in the technology itself, but also in the understanding of user needs, the speed of product polishing, and the resulting user habits and brand effect.
As you can see from our review and testing, Manus is trying to build its own barriers in the latter way. Including its intensive meeting with users around the world, it is also part of this strategy.
But it is destined to be a race against time. The competition in the AI Agent track has become extremely fierce. With giants such as OpenAI and Google around, and countless flexible startups chasing after it, whether Manus can rely on faster iteration speed to build enough user loyalty before the underlying technology catches up will be the key to determining its fate.
Judging from these product experiences, the product has not stopped updating, iterating and improving functions in the process of “running away”, but at the same time, these functions have not completely distanced it from other opponents. If domestic and foreign teams, operations and other affairs can take up less energy, and more investment in product and technology iteration, what can the product become of Manus? This is always the real question.
