Some exclusive information and slow thinking about Manus

1. We learned about the Manus project at a relatively early stage. The reason why we did not rush to publish the article like other media after its release last night is that we have been exposed to and used many similar frameworks before. Before we have real experience, it is difficult for us to feel its uniqueness, and it is difficult for us to imagine how much progress the update of non-underlying models can bring. Therefore, we did not make any public .

2. We are very honored and grateful that the Manus team gave us the opportunity to participate in the closed-door communication meeting this morning and experience the product first. The Manus team believes that this is "The next ChatGPT moment". Everyone in the team and the circle is enthusiastic and the atmosphere is very good.

3. Manus was previously an AI browser, and then the Arc team transformed to Dia and encountered similar problems, but they did more and faster than them, so they switched to the current Manus. The previous human-computer collaboration experience of Browse Use and Computer Use was not good. AI was competing with users for control. After you gave a task, you could only watch the performance of AI from the side. If you touched it by mistake, the process might be interrupted. AI needs to use a browser, but the Manus team believes that AI should be given its own cloud browser, and finally the results should be fed back to the user.

4.  Less Structure , More Intelligence.  This is a non-consensus that is discussed a lot in the industry. The heated discussion on this issue has been going on since platforms such as Coze supported building AI applications through workflow. For example, Flood Sung said when Kimi released k1.5, "The current various Agentic Workflows are all things with Structure, which will definitely limit the model's 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 workflow. All capabilities are naturally evolved from the model, rather than taught by workflow.

5. The first core of Manus design is to equip the big model with a computer , so that it can plan step by step to do observation and action; the second core is to equip it with system permissions, just like opening some company account permissions for new colleagues. Manus is connected to a large number of private APIs and can process a lot of structured authoritative data; the third core is to train it, training it is like having a running-in process with new colleagues. Manus will also continue to learn your requirements based on your usage habits.

6. The Manus team also invented a very interesting indicator, which is to measure the success of current AI products. Instead of looking at DAU, they look at  AHPU , Agentic Hours Per User, which is an indicator that measures how much agent runtime a single user uses, and how many agent hours a user calls every day.

7. The Manus model uses Claude and Qwen, and the cost of a single task is about two dollars.

8. As for why the Manus team created this product? Manus replied that it was the right time, right place, and right people, as well as the team's implementation of the "Less Structure, More Intelliengence" concept. The agility and decision-making speed of the startup team outperformed the overly structured organizational inertia of large companies.

9. Why does Manus say that it is "the world's first truly universal AI Agent?" Do the previous Operator, Deep Research, MetaGPT, AutoGPT, Eko, etc. not count? And why do some people say that Manus is a shell to the extreme? In our understanding, some previous Agent open source frameworks can actually achieve similar effects to Manus, but Manus has made some good engineering optimizations and was the first to be commercialized. Here I would like to cue another Chinese team  Flowith . The Oracle model they made half a year ago  can basically achieve the effects demonstrated by the current Manus Demo.

10. Why is Manus so popular, even though 99% of people cannot experience it now? First of all, the Manus team itself is very good. They and Zhen Fund have a very good reputation and influence in the circle, and many KOLs are willing to spread the word for them. Second, the strategic setting of the invitation code is really fissionable. People in China are really used to and enjoy "crazy grabbing" and "hype", and they can even "do everything possible", which has been reflected in many previous events.

11. People always go from one extreme to another. Our attitude is not to praise or diss too much, and even many conspiracy theories have emerged...too much is as bad as too little. We should be able to look at Manus and the Chinese entrepreneurial team with enough tolerance and rationality. The current popularity is also beyond the expectations of the Manus team. The reason for setting up an invitation code is that Multi-Agent consumes a lot of computing resources, and it is difficult for such an entrepreneurial team to fully open the product to all users.

12. Today, too many friends have come to us to ask for invitation codes and what they think of Manus. In our opinion, we are not willing to draw conclusions too early for traffic. Giving emotional and overly favorable reviews is a kind of harm to the product itself and the AI ​​ecosystem. Readers are quick to get into it and quick to leave it, which is easy to cause aesthetic fatigue and trust crisis.

13. After a short test with our internal test partners, we found that Manus still has some illusions and inevitably has some corner cases, which is still a long way from being truly universal. Please give us some more time for specific experience and evaluation.

14. We are very happy to see more and more people learn about AI Agent because of Manus. As one of the earliest teams in China to start researching and developing Agent, Agent Universe will bring you more related reviews and reports in the future.