Manus, the AI general agent launched just last week on the internet, has taken off like wildfire. It’s not only in China where the Wuhan startup Butterfly Effect developed it. The performance of this AI model has been praised by influential voices, such as Twitter cofounder Jack Dorsey, Hugging Face’s product leader Victor Mustar and Hugging Face CEO Victor Mustar. Many have dubbed this “the second DeepSeek,” as it is compared to an earlier AI model which surprised the industry with its surprising capabilities and origin.
Manus is the first AI general agent. It uses multiple AI models, such as Anthropic’s Claude 3.5 Sonnet or finely tuned versions of Alibaba’s Qwen open-source software to perform a variety of autonomous tasks. This makes it a different AI chatbot, such as DeepSeek which is based on one large language family and designed primarily for conversational interaction.
In spite of all the buzz, only a few have actually used it. Under 1% of users who are on the waiting list currently have an invitation code. It’s not clear how many users are currently on the list. However, Manus Discord has over 186,000 subscribers, which gives you an idea of interest.
MIT Technology Review Manus was available to me and I took it for a spin. It’s remarkably flexible, explains itself clearly and is very adaptable. It’s promising, but it isn’t perfect.
Manus, like the AI assistant Monica, released by its parent company in 2023 and aimed at a worldwide audience, is also marketed to a global market. The default language is English, and the design is minimalist and clean.
A valid invitation code is required to gain access. After that, the system will direct users to an entry page similar in design to those found on ChatGPT and DeepSeek. Previous sessions are displayed along the left column with a chat box at the center. This landing page includes sample tasks that have been curated by the firm, ranging from interactive learning sessions to audio meditations.
Manus, like other AI agents that use reasoning, such as ChatGPT, can break down tasks into smaller steps, and navigate the internet to find the necessary information to finish them. It is distinguished by the “Manus’s Computer”, which lets users observe the agent’s actions and intervene at any time.
Manus was given three tasks to test his abilities: 1) compile a listing of reporters who cover China technology, 2) search New York City for listings for properties with two bedrooms, and 3) nominate candidates for the Innovators under 35 list, created by MIT Technology Review Every year,
How to do it:
Step 1: Manus’ first list, which I received from him, contained five reporters with “honorable mentions” listed below. It listed the notable works of some journalists, but not others. Manus asked me why. I asked Manus why. It gave me a hilariously simple answer: it got lazy. The agent explained that it was partly due to the time constraint as she tried to speed up the research. Manus provided a list of thirty journalists who were noted for their work and current publication. It was a pleasure to be included in the list, alongside many of my friends and colleagues.
It was impressive that the program responded to my suggestions, just as I would to a real intern or assistant. It initially ignored changes to the employer status of some journalists, but when I requested it revisit certain results, it corrected them quickly. It was also nice that the output could be downloaded as either a Word file or Excel document, which made it easier to share or edit.
Manus ran into a problem when trying to access news articles from journalists behind paywalls. It frequently came across captchas. As I followed along with the steps, it was easy to take control and complete them, even though some media websites still block this tool on suspicion of suspicious behavior. There is a lot of room for improvement here. It would be helpful if future versions of Manus asked for assistance when they encountered these restrictions.
Step 2: Manus was given a set of complex criteria for the apartment search. These included a budget, several parameters, and access to Manhattan downtown. I also asked him to find an apartment that had a large kitchen and outdoor space. Manus took vague criteria like “some sort of outdoor space”, too literally at first, and excluded properties that did not have a balcony or private terrace. After more clarification and guidance, Manus was able compile a more comprehensive and helpful list. It included recommendations arranged in neat bullet points and tiers.
It was a direct feeling from the final product. WirecutterThe online property listings were more easily accessible and structured.
Step 3: Manus was asked to submit 50 names for the Innovators Under 35 List. This list takes a lot of work to produce, and every year we receive hundreds of nominations. Manus was my first choice to test. The task was broken down into several steps. These included reviewing previous lists in order to better understand the selection criteria.
Manus spent the longest time developing a strategy for searching. The Manus Computer’s window did not explicitly state its strategy, but it showed the agent quickly scrolling through the websites of research universities and announcements about tech awards. It encountered problems again when it tried to access media and academic content that was behind paywalls.
It was only after three hours spent scouring through the Internet, during which Manus repeatedly asked me if I wanted to narrow my search that it could provide me with three full profiles of candidates. It eventually produced a list, though certain fields and academic institutions were overrepresented. This was a result of an incomplete search process. It was only after I raised the problem and requested that it find five Chinese candidates, did it manage to produce a list of five names. However, the results were skewed towards Chinese media favorites. The system told me that Manus might suffer if I continued to input too much text.
What I think: Overall, Manus is a very intuitive and user-friendly tool that can be used by users who have or do not have coding experience. It produced better results on two out of three tasks than ChatGPT DeepResearch despite taking significantly more time to finish. Manus is best for analytical tasks that involve extensive internet research but are limited in scope. It’s better to limit your tasks to those that a human expert could perform in a single day.
It’s still not smooth sailing. Manus may experience frequent crashes, system instability and difficulty processing large amounts of text. Due to high load on the service, new tasks are not possible. When I attempted to create new requests on the computer, the message “Please try again in few minutes” appeared on my screen. Sometimes Manus’s Computer would freeze on one page for an extended period.
Peak Ji, Manus’s Chief Scientist, says that it has a greater failure rate compared to ChatGPT DeepResearch. This is something the team is working on. The Chinese media outlet has said that the team is working to address this problem. 36Kr Reports show that Manus costs $2 per task, just one tenth the cost of DeepResearch. The Manus team could make the software a favorite among individual users. This would include white-collar workers, developers and small teams.
It’s also important that the working process of Manus feels transparent and collaborative. The agent asks for questions and stores key instructions in memory as “knowledge”. This allows for a highly customizable experience. Each session can be replayed and shared, which is a nice feature.
Manus will continue to be a part of my life, both personally and professionally. Although I don’t think the DeepSeek comparisons are accurate, this is further proof that Chinese AI firms aren’t just copying their Western counterparts. They are not just modifying base models; they actively shape the adoption of AI agents.
