Since Manus, the general AI agent, was launched last Thursday, it has spread like wildfire online. It’s not just in China where it was developed by Wuhan-based startup Butterfly Effect. It has made its way to the global conversation. Influential voices in tech such as Twitter cofounder Jack Dorsey, and Hugging Face’s product lead Victor Mustar have praised its performance. Some have even dubbed this AI model “the second DeepSeek,” as it compares it to the earlier AI that surprised the industry with its unexpected capabilities and its origin.
Manus is the world’s “first general AI agent”using multiple AI models, such as Anthropic’s Claude 3.5 Sonnet or fine-tuned Qwen versions from Alibaba. It can perform a variety of tasks autonomously. This makes it different from AI bots, such as DeepSeek, that are based on one large language model and are designed primarily for conversational interaction.
Despite the hype, only a few people have used it. Currently, less than 1% of users on the waiting list have received an invitation code. It’s not clear how many people are waiting on the list, but Manus’s Discord has over 186,000 members.
MIT Technology Review (19459010) was able obtain access to Manus. When I gave it a try, I found it to be like working with a highly intelligent intern. It can sometimes lack understanding of what’s being asked, make incorrect assumptions, or cut corners to expedite a task. However, it explains it’s reasoning clearly, is remarkably adaptive, and can significantly improve when given detailed instructions or feedback. It’s promising, but not perfect. Manus, like the AI assistant Monica, released by its parent company in 2023, is aimed at a global audience. The default language is English, and the design is minimalist and clean.
A user must enter a valid invitation code to gain access. The system then directs users to an landing page that closely mimics those of ChatGPT and DeepSeek. Previous sessions are displayed in the left-hand column, while a chat input field is located in the middle. The landing page includes sample tasks that have been curated by the firm, ranging from business strategy to interactive learning and customized audio meditation sessions.
Manus, like other reasoning-based AI tools such as ChatGPT DeepResearch is capable of breaking down tasks into steps and autonomously browsing the web to find the information it requires to complete them. The “Manus’s Computer window” allows users to not only observe what the agent does, but also to intervene.
I put Manus to the test by giving him three tasks: (1) compile an extensive list of reporters who cover China tech; (2) search for listings for two-bedroom properties in New York City; and (3) nominate candidates for Every year, MIT Technology Review creates a list of Innovators Under35.
Task 1 Manus gave me a list of five reporters, and five “honorable” mentions below. I noticed it listed some journalists notable work, but not others. I asked Manus what the reason was. The answer was hilariously simple. It got lazy. The agent explained that it was “partly because of time constraints, as I tried expediting the research process.” Manus responded to my request for consistency and thoroughness with a list of 30 journalists who were noted by their current publication and notable work. (I was happy to see that I made the cut along with many of my favorite peers.)
It was impressive that I could make suggestions at the highest level, just as I would for a real intern or assistant. And that it responded accordingly. It initially ignored changes in the employer status of some journalists, but when I asked it revisit some results, they were quickly corrected. The output is also downloadable as a Word file or Excel document, making it easier to edit or share.
Manus encountered a problem, however, when trying to access journalists’ news articles that were behind paywalls. It frequently encountered captcha blockers. I could easily complete these because I was able follow along step-by-step, even though many media sites blocked the tool citing suspicious activities. I see major improvements in this area–and a future version Manus should be able to ask for assistance when it encounters such restrictions.
I gave Manus an extensive set of criteria for the apartment search. These included a budget, several parameters, and a large kitchen. Manus took vague requirements such as “some sort of outdoor space” literally at first, excluding all properties without a balcony or private terrace. After more clarification and guidance, the list was able compile a more comprehensive and helpful list with neat bullet points and tiers.
This final output was straight out of Wirecutter, with subtitles such as “best overall,” ‘best value,” and ‘luxury option. It took less than a half-hour to complete this task.
The third task was the most ambitious: I asked Manus for 50 nominations for this year’s Innovators Under 35. This list is a huge undertaking and we receive hundreds of nominations each year. So I was curious to find out how well Manus would do. It broke down the task into several steps, such as reviewing past lists to better understand the selection criteria, creating an effective search strategy to identify candidates, compiling names and ensuring that a diverse selection from around the world was made.
The most time-consuming task for Manus was developing a search strategy. The Manus’s Computer window, which didn’t specify its approach, showed the agent quickly scrolling through websites for prestigious research institutions, announcements of technology awards, and news articles. It encountered problems again when trying to access academic articles and paywalled content.
Despite spending three hours scouring the web, during which Manus repeatedly asked me if I wanted to narrow my search, it was only able provide me with three candidates who had full background profiles. It eventually produced a list of 50 names when I asked for it, but certain institutions and fields were overrepresented. This was a result of an incomplete research process. After I raised the issue and asked for five candidates from China, the system managed to compile an impressive five-name list. However, the results were heavily skewed towards Chinese media darlings. I finally gave up when the system warned me that Manus might suffer if I continued to input too much text.
Overall Manus is a very intuitive tool that can be used by users with or without programming backgrounds. It produced better results on two of the three tasks than ChatGPT DeepResearch despite taking significantly longer to complete. Manus is best suited for analytical tasks that require extensive internet research but are limited in scope. It’s best to limit your tasks to those that a human intern would be able to do in a single day. Manus may experience frequent crashes and system instabilities, and it might struggle to process large amounts of text. Due to the high service load at this time, tasks cannot be created. When I tried to create new requests, the message “Please try again in a couple of minutes” flashed a few times on my screen. And occasionally Manus’s Computer would freeze on a particular page for a prolonged period of time. Peak Ji, Manus’s chief scientific officer, told that the team is working to address this problem.
The failure rate of ChatGPT DeepResearch is higher. The Chinese media outlet 36kr reported that Manus costs $2 per task, which is only one-tenth the cost of DeepResearch. If the Manus team improves its server infrastructure I can see it becoming a favorite tool for individuals, especially white-collar professionals and independent developers.
Lastly, I think that it’s really important that Manus’s work process feels relatively collaborative and transparent. It actively asks you questions along the way, and stores key instructions in its memory as “knowledge”. This allows for a highly customizable agentic experience. It’s nice that you can replay and share each session.
In my personal and professional life, I’m sure I will continue to use Manus for a variety of tasks. Although I’m not certain the comparisons with DeepSeek are accurate, it serves to further prove that Chinese AI companies do not simply follow in the footsteps their Western counterparts. They are not just modifying base models; they are actively shaping AI agents to their own needs.
