Everyone in AI is talking about Manus. We put it to the test.

Since the launch of the Manus General AI agent last week, it has spread online like the wilderness. And not only in China, where it was developed by the starting starting butterfly impact in Wuhan. He made his way to global conversation, with influential technology sounds, including the founder of Twitter Jack Dormyy and Huging Face Victor Mustar, praising his performance. Some even called it “The Second Deepseek”, compared to the previous artificial intelligence model that drew the amazement of the industry with its unexpected capabilities as well as its origin.

Manus claims to be the first public agent in the world, Amnesty International, as he takes advantage of the Multiple Models of Amnesty International (such as Claude 3.5 from Anthropor, the well -seized versions of QWEN Open Source in Alibaba) and many independent working factors to independently act on a wide range of tasks. (This makes it different from AI Chatbots, including Deepseek, which depends on one large model family mainly designed for conversation interactions.)

Despite all this noise, a few people have an opportunity to use it. Currently, less than 1 % of users in the waiting menu receive an invitation code. (It is not clear that the number of people in this list, but for the amount of interest, the Manus Discord channel has more than 186,000 members.)

Massachusetts Institute Technology Review Technology I managed to reach Manus, and when I gave him a test, I found that his use seems to be cooperating with a very smart and effective trainee: while he sometimes lacks understanding what he is asked to do, and makes incorrect assumptions, or cuts angles or discounts in expediting tasks, she explains clearly, and can improve significantly, and can significantly improve with detailed instructions. Ultimately, it’s promising but not perfect.

Just like the former producer of its mother company, an Amnesty International Assistant called Monica was released in 2023, Manus is intended for a global audience. The English language was set as a virtual language, and its design is clean and a minimum.

To enter, the user must enter a valid invitation code. Then the system directs users to a deliberately similar page to the ChatGPT or Deepseek, with the previous sessions displayed in a left column and a chat entry box in the middle. The intended page also features sample tasks sponsored by the company – starting from developing the business strategy to interactive learning to dedicated audio meditation sessions.

Like thinking tools based on thinking, such as ChatGPT Deepresearch, Manus is able to divide tasks into steps and move the web independently to obtain the information he needs to complete. What distinguishes it is the “Manus’s Computer” window, which allows users not only to monitor what the agent does but also to intervene at any time.

To put it in the test, Manus gave three tasks: (1) Collecting a list of prominent correspondents covering China Tech, (2) Look for two bedroom lists consisting of two bedrooms in New York City, and (3) nomination of potential candidates for innovators under the age of 35, a list created by Massachusetts Institute Technology Review Technology every year.

Here is how you did:

Mission 1: The list of the first reporters that Manos Lee gave me only five names, with five “honorable signals” under it. I noticed that he included the work of some prominent journalists, but they did not do it for others. I asked Manus why. The reason he presented was very simple: he became lazy. The client told me: “This was partly due to the restrictions of time as I tried to accelerate the search.” When she insisted on consistency and accuracy, Manos responded to a comprehensive list of 30 journalists, referring to their current outlet and the inclusion of prominent work. (I was happy to see that I cut, along with many of my beloved peers.)

I was affected that I was able to make suggestions from the highest level of changes, as much as someone does with a realistic trainee or assistant, and that he had responded appropriately. While she was initially overlooked in the case of the employer of some journalists, when she asked him to reconsider some of the results, she quickly corrected them. Another nice feature: the output is downloaded as a word or an Excel file, which makes it easy to adjust or share it with others.

Manus struck an obstacle, despite reaching news articles for reporters behind Paywalls; Often faced captcha blocks. Since I managed to follow along step -by -step, I was able to easily complete it, although many media sites still prevent the tool, indicating suspicious activity. I see the possibility of significant improvements here – and it will be useful if I can release my future from Manus proactively when he faces these types of restrictions.

Mission 2: For research in the apartment, Manus has given a complex set of standards, including the budget and many parameters: a wide kitchen, an outdoor space, access to downtown Manhattan, and a main train station seven minutes. Manus initially interpreted mysterious requirements such as a “type of external space” very literally, with the exception of real estate without reaching a private balcony or balcony. However, after further guidance and clarification, I managed to collect a broader and more useful menu, while making recommendations in the elegant levels and lead points.

Final product hair directly from WIRECTERThat contains translations such as “Best Compene”, “Best Value”, and “Luxury Option”. This task (including face and back) took less than half an hour-less than a lesser time than assembling the list of journalists (which lasted slightly more than an hour), most likely that real estate lists are more explicitly available and online.

Mission 3: This was the largest in the scope: I asked Manus to nominate 50 people for this year’s innovators under 35 lists. The production of this list is an enormous task, and we usually get hundreds of nominations every year. So I was curious to see the quality of Manus’s work. It has set up the task to steps, including reviewing the previous lists to understand the selection criteria, creating a research strategy to identify candidates, collect names, and ensure a variety of candidates from all over the world.

The development of the research strategy was the most consuming time for Manus. Although it has not explicitly specified its approach, the Manus computer window revealed the rapid pass through web sites of prestigious research universities, technical awards and news articles. However, she again faced obstacles when trying to reach academic papers and restricted media content.

Three hours after the Internet – that is, Manus (from the concept) asked me several times if I could narrow the search scope – was able to give me only three candidates with full rear profiles. When I pressed him again to provide a full list of 50 names, she finally created one, but some academic institutions and fields were greatly represented, reflecting an incomplete search process. After I referred to this issue and asked to find five candidates from China, I managed to collect a solid list of five names, although the results were heading towards the columns of Chinese media. In the end, I had to surrender after the regime warned that Manus’s performance may decrease if it continued to enter a lot of text.

My evaluation: total, I found that Manus is a very intuitive tool for users who have or without coding wallpapers. In two three tasks, better results than Chatgpt Deepresearch were provided, although it took much longer to complete it. Manus is the most appropriate for analytical tasks that require intense search on open internet but has a limited range. In other words, it is better to stick to the types of things that a skilled human trainee can do during a day of work.

However, it is not everything smooth. Manus can suffer from repeated incidents and instability in the system, and he may fight when he is asked to treat large parts of the text. The message “Because of the current high service download, the tasks cannot be created. Please try again within a few minutes on my screen several times when I tried to start new requests, and sometimes the Manus computer freezed on a specific page for a long period of time.

It has a higher failure of Chatgpt Deepresearch – a problem that the team treats, according to the chief scientist in Manus, PEAK JI. However, the Chinese media 36 each Reports that Manus costs all tasks is about $ 2, which is just the cost of Deepresearch. If the Manus team enhances its server infrastructure, I can see the tool that becomes a favorite choice for individual users, especially professionals with white collars, independent developers and small teams.

Finally, I think it is really important to be the work of Manus’s transparent and relatively cooperation. He actively asks questions along the way and maintains the main guidelines as “knowledge” in its memory for future use, allowing an easily customized agent experience. It is also good for each session to be shared and shared.

I expect to continue to use Manus for all types of tasks, in both my personal and professional life. Although I am not sure that the comparisons with Deepseek are completely correct, they work as an additional evidence that Chinese artificial intelligence companies are not only followed in the footsteps of their Western counterparts. Instead of just innovation in basic models, it is actively the adoption of artificial intelligence agents in their own way.

2025-03-11 22:21:00