Source: Geek Park
Author | Shiyun Zhang Yongyi
The year 2025 is the inaugural year of AI Agents—this statement was confirmed in the early morning of March 6, Beijing time.
After DeepSeek, another sleepless night in the tech circle.
Many users commented like this on Social Media.
Everyone stayed up all night for a single usage invitation code for this product - it is the world's first AI Agent product "Manus" developed by Monica.im.
According to the team, "Manus" is a truly autonomous AI agent capable of solving various complex and changing tasks. Unlike traditional AI assistants, Manus not only provides suggestions or answers but can also directly deliver complete task results.
As the name 'Manus' suggests, it symbolizes 'hand' in Latin. This means that knowledge should not only be in the mind but also be executable by hand. This is the essential advancement of Agent and AI Bot (chatbot) products.
Where are Manus's cows? The most direct way is to look at the official website display and the user self-developed use cases, part of which are organized as follows by Geek Park:
Travel Planning: Not only integrates travel information but also creates customized travel manuals for users. For example, planning a trip to Japan in April, providing personalized travel advice and detailed manuals.
Stock Analysis: Conduct in-depth stock analysis and design visually appealing dashboards that display comprehensive stock insights. For example, performing in-depth analysis on Tesla stocks and creating a visual dashboard.
Educational Content Creation: Create video presentation materials for middle school teachers, explaining complex concepts like the momentum theorem, helping teachers teach more effectively.
Insurance Policy Comparison: Create clear insurance policy comparison charts, provide the best decision-making advice, and help users choose the most suitable insurance products.
Supplier Procurement: Conduct in-depth research across the entire network to find suppliers that best meet user needs, serving users as a truly fair agent.
Financial Report Analysis: Capture market sentiment changes towards specific companies (such as Amazon) through research and data analysis, providing market sentiment analysis for the past four quarters.
Sorting the list of Ventures: Visit relevant websites to identify companies that meet the criteria, and organize them into a table. For example, compile a list of all B2B companies from YC W25 batch.
Online store operation analysis: Analyze Amazon store sales data, provide actionable insights, detailed visualizations, and customized strategies to help improve sales performance.
When the Agent produces an incredibly complete and professional result through a long chain of thinking and tool calls, users begin to exclaim, "It really can help humans get things done."
According to official website information, in the GAIA benchmark test (which assesses the ability of general AI assistants to solve real-world problems), Manus achieved new state-of-the-art (SOTA) performance at all three difficulty levels.
In summary—what Manus really wants to do is to be your literal "agent" in the digital world. And it has achieved that.
Just as you imagined, Manus, launched at midnight, suddenly woke up everyone in the AI community!
Manus, your "digital agent"
First of all, the biggest difference in experience with Manus compared to previous LLMs is:
It emphasizes the ability to deliver final results directly, rather than just providing a simple "answer."
Currently, Manus adopts a Multiple Agent architecture, operating in a manner similar to the Computer Use model previously released by Anthropic, fully running in independent virtual machines. It can also call various tools in virtual environments — writing and executing code, browsing the web, operating applications, etc., directly delivering complete results.
In the official release video, three work cases completed by Manus in practical use scenarios are introduced:
The first task is to screen resumes.
From 15 resumes, recommend suitable candidates for the reinforcement learning algorithm engineer position, and rank the candidates based on their knowledge of reinforcement learning.
In this demonstration, there is no need to extract compressed files or manually upload each resume file one by one. Manus has already shown itself to be like a human "intern," manually extracting files, browsing each resume page by page, while recording important information.
In the results provided by Manus, there are not only automatically generated ranking suggestions, but it also classifies candidates into different levels based on important dimensions such as work experience. Upon receiving the user's preference to present the information in an Excel spreadsheet, Manus can automatically generate the corresponding table by writing a Python script on-site.
Manus can also record information such as "users prefer to receive results in a tabular format" through memory capabilities during this practice process, and will prioritize presenting results in a table the next time similar tasks are handled.
The second case, tailored for the local people, is selecting properties.
In the case, the user wishes to purchase property in New York, with input requirements for a safe community environment, low crime rates, and quality primary and secondary education resources—of course, including the most important budget that is affordable on a fixed monthly income.
In this demand, Manus AI breaks down complex tasks into a to-do list, including researching safe communities, identifying quality schools, calculating budgets, and searching for properties. It carefully reads articles about New York's safest communities through online searches, collecting relevant information.
Secondly, Manus writes a Python program to calculate the affordable property budget based on user income. It filters the property list based on budget through relevant housing price information on real estate websites.
Finally, Manus integrates all collected information and writes a detailed report, including community safety analysis, school quality assessment, budget analysis, recommended property list, and relevant resource links—like a professional real estate agent. Moreover, due to Manus’s attribute of being "completely user-interest oriented", its use and experience are even better.
In the last case, Manus demonstrated its ability to analyze stock prices.
The task given in the case was to analyze the correlation between the stock prices of NVIDIA, Marvell Technology, and Taiwan Semiconductor over the past three years: it is well known that there is a close correlation among these three stocks, but for new users, it is difficult to quickly sort out the causal relationships.
Manus's operation is very similar to that of a real stockbroker. It first accesses information websites like Yahoo Finance through an API to obtain historical stock data while cross-validating the accuracy of the data to avoid being misled by a single source of information, which could significantly impact the final results.
In this case, Manus also utilized its ability to write Python code, conduct data analysis, and perform visualization, while also introducing professional financial tools for analysis. Ultimately, it provided users with causal relationships through data visualization charts, paired with a detailed comprehensive analysis report — truly resembling the daily tasks of an intern in the financial field.
Moreover, the official Manus website showcases more than ten scenarios where Manus can be used: directly using Manus to organize your itinerary, personalized travel route recommendations, and even allowing it to learn to use various complex tools to streamline everyday tasks.
In this process, what truly sets Manus apart from traditional tools is its self-planning capability to ensure task execution.
The self-learning ability also makes Manus's work capability enhancement logic more akin to that of real humans — even at this stage, it may not be able to reach expert-level proficiency in any specific field, but a huge potential can already be seen.
With the addition of self-learning capabilities, the versatility of the AI Agent has greatly increased. In actual tests conducted by users on Manus, you can even directly describe relevant content from a video frame to it, and Manus would ultimately be able to accurately find the link to a specific Douyin Short Video across platform content limitations based on the corresponding information.
Since the current version of Manus operates entirely in the cloud and asynchronously, its capabilities are not actually limited by factors like the end-side platform form or computing power you use—users can even temporarily shut down their computers after issuing commands to Manus, and it will automatically notify you of the results once the activity is complete.
This operational logic is also very familiar — it's like a person after work shouting to an intern on WeChat to 'send me the organized files.' The only difference is that now, this intern can truly respond to you 24/7, without worrying about 'office discipline.'
Multiple agents + self-inspection, running through the AI Agent flow.
From the above cases, it is not difficult to see that Manus's real ace is not the 'AI Agent' concept that has already appeared in Computer Use, but its ability to 'work in a human-like way.'
Compared to 'running calculations,' Manus's work logic is more like 'thinking and executing commands.' It does not accomplish things that humans cannot currently do, which is why some users who have experienced the current version of Manus describe it as 'an intern.'
On the Manus official website, there are many tasks that Manus can accomplish, including a case that demonstrates how to use Manus in B2B Business. It quickly and accurately matches your ordering needs with global suppliers.
In the conventional products with similar demands, integrating global supply chain enterprise information within the platform to help users complete supplier/demand side matching is a common industry logic. However, in Manus's case, you can see a completely different implementation method.
Manus AI uses a framework called 'Multiple Agent,' running in independent virtual machines. By coordinating the efforts of planning agents, executing agents, and validating agents, it significantly improves the efficiency of handling complex tasks and shortens response times through parallel computing.
In this architecture, each agent may be based on independent language models or reinforcement learning models, communicating with each other through APIs or message queues. At the same time, each task runs in a sandbox, avoiding interference with other tasks, while supporting cloud expansion. Each independent model can mimic the way humans handle tasks, such as thinking and planning first, understanding complex instructions, breaking them down into executable steps, and then calling the appropriate tools.
In other words, through this multi-agent architecture of Manus, it is more like being assisted by multiple assistants, completing tasks such as resource retrieval, docking, and validating the effectiveness of information to help you with the entire workflow—this actually resembles not just hiring an 'intern', but directly becoming a miniature 'department head'.
In the case of B2B Business, Manus uses web crawlers along with code writing and execution capabilities to automatically search in the vast sea of the Internet, matching potential suppliers based on your own needs in terms of product quality, price, delivery capabilities, etc. It can not only present conclusions in a graphical manner right before you but also provide more detailed operational suggestions on this data.
As for how the Monica team achieved the video effects and what technology was used, reports suggest the team will reveal this on March 6th, Beijing time.
The ultimate of 'sewing' is to explode.
What kind of company is Monica.im, the force behind Manus?
Monica is an All-in-One AI assistant, with its product form slowly expanding from a browser plugin to an app and web version. The mainstream use case is when users click its small icon in the browser, they can directly use the major mainstream models it connects to. By accurately understanding the user needs in segmented scenarios, Monica has picked the 'low-hanging fruit' of large models.
The founder, Xiao Hong (nickname Little Red, English name Red), is a young serial entrepreneur, born in 1992 and graduated from Huazhong University of Science and Technology. After graduating in 2015, he started his own business, which did not go smoothly in the early stages (such as campus social networking and second-hand markets). In 2016, he created a tool for WeChat public account operators for editing and data analysis, gaining a million users and achieving profitability, and the final product was sold to a unicorn company in 2020.
By 2022, with the wave of large models, he officially founded Monica, focusing on the overseas market, and quickly achieved a cold startup with the independently developed product ChatGPT for Google.
In 2024, as soon as the GPT-4o, Claude 3.5, and OpenAI o1 series were launched, Monica enabled users to access the latest SOTA models. With new advancements in connected models, the professional search, DIY Bot, Artifacts write mini-programs, and memory functions launched by Monica also gained popularity among users. Monica presents different interactive forms and functionalities across various web pages like YouTube, Twitter, Gmail, and The Information to adapt to specific user needs in certain scenarios, updating the personalized AI experience of hundreds of web pages.
In 2024, the number of Monica users doubled to 10 million. At the same time, it maintained considerable profitability, ranking at the top among similar overseas products.
Monica's strong performance validates one thing:
Perfectly following the shell strategy leads to both TPF and PMF, ultimately leading to user value.
Manus may continue this thought of the Monica team — Xiao Hong stated in an interview with media personality Zhang Xiaojun that products cannot only have one form of chatbot; agents will be a new form that requires new products to accommodate.
He drew inspiration from the AI programming products Cursor and Devin. According to Geek Park, the former mainly operates in copilot mode, while the latter is in autopilot mode, which better meets human needs. Agent should also be aimed at the general public, truly executed by AI, just like Devin. However, the past issue was that the model was not intelligent enough.
However, based on the existing capabilities of the model to package services for scenarios, this may indeed be the advantage of the Monica team. Xiao Hong mentioned that currently there are not many Agent product teams because it requires various complex abilities. For instance, the team needs experience with chatbots, AI programming, and browser-related functions (since everything runs in the browser), and there must be a good perception of the model's boundaries—what level it has developed to today and what levels it will reach in the future, etc.
"There are not many companies that possess all these capabilities, and those that do may be engaged in a very specific business, but we happen to have team members who have the time to work together to accomplish this task." he said.
When summarizing why Monica was able to accomplish this, he said, "Firstly, I think we are quite lucky. Secondly, to some extent, if everyone starts focusing on reasoning today, could it be that more time is actually freed up for Ventures? How far can the model's expected capabilities overflow?"
He believes that the Agent is still in its early stages. First, the Agent is still in the planning phase and hasn't yet moved to execution in the physical world; second, the capabilities of large models are still on the rise, and everything remains unpredictable.
"I certainly do not know that an Agent can be produced in this way; it is an unknown matter," he said.
Interestingly, Monica, who "doesn't know how to make an Agent," has now created a product that has shocked the entire AI community.
Manus may not be the ultimate AI Agent, but it has undoubtedly raised people's expectations of AI by another magnitude after the explosive success of DeeoSeek.
Editor/ping
