2025 is the first year of the AI Agent - this statement was verified in the early morning of March 6th, Beijing time.

"After DeepSeek, another sleepless night in the tech circle."

Many users commented on social media like this.

Everyone stayed up all night, just for an invitation code to use the 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 that can solve various complex and changing tasks. Unlike traditional AI assistants, Manus not only can provide suggestions or answers, but also can directly deliver complete task results.

The introduction video of Manus is only 4 minutes long, but it is very powerful | Image source: Monica.im

As the name "Manus" implies, it symbolizes "hand" in Latin. That is to say, knowledge not only needs to be in the brain, but also needs to be executable by hand. This is the essential evolution of Agent and AI Bot (chatbot) products.

Where is Manus outstanding? The most intuitive is to look at the official website display and the user-generated use cases, which Geek Park has partially summarized as follows:

Travel planning: Not only integrating travel information, but also creating customized travel manuals for users. For example, planning a trip to Japan in April, providing personalized travel recommendations and detailed manuals for users.

Stock analysis: Conducting in-depth stock analysis, designing visually appealing dashboards to display comprehensive stock insights. For example, conducting in-depth analysis of Tesla stock and creating a visualization dashboard.

Educational content creation: Creating video demonstration materials for middle school teachers to explain complex concepts like the momentum theorem, helping teachers teach more effectively.

Insurance policy comparison: Creating clear insurance policy comparison tables, providing the best decision recommendations, and helping users choose the most suitable insurance products.

Supplier procurement: Conducting in-depth research across the entire network, finding the most suitable suppliers for user needs, and serving as a truly fair agent for users.

Financial report analysis: Capturing market sentiment changes for specific companies (such as Amazon) through research and data analysis, providing market sentiment analysis for the past four quarters.

Startup company list organization: Visiting relevant websites to identify qualified companies and organizing them into a table. For example, compiling a list of all B2B companies in the YC W25 batch.

Online store operations analysis: Analyzing Amazon store sales data, providing actionable insights, detailed visualizations, and customized strategies to help improve sales performance.

When the Agent outputs an extremely complete and professional result through a long chain of thinking and tool invocation, users begin to exclaim "It can really help humans do things".

According to the official website information, in the GAIA benchmark test (evaluating the ability of general AI assistants to solve real-world problems), Manus has achieved new state-of-the-art (SOTA) performance at all three difficulty levels.

In summary - Manus wants to be your "agent" in the digital world, and it has achieved that.

Just as you thought, the launch of Manus in the early morning has awakened all the people in the AI circle!

01 Manus, your "digital agent"

First, the biggest difference between Manus and previous LLMs in terms of experience:

It emphasizes the ability to directly deliver the final result, rather than just giving a simple "answer".

Manus currently uses a Multiple Agent architecture, with a running mode similar to the Computer Use released by Anthropic, running completely in an independent virtual machine. It can also call various tools in the virtual environment - writing and executing code, browsing web pages, operating applications, etc., to directly deliver complete results.

In the official release video, three cases of Manus' work in actual use scenarios were 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 reinforcement learning expertise.

In this demonstration, you don't even need to unzip the compressed file and manually upload each resume file. Manus has already shown the "intern" side like a human, manually unzipping the file and browsing each resume page, while recording the important information.

Manus, like an intern, automatically understands the hidden instruction "unzip the packed file the boss threw over" | Image source: Geek Park

In the results provided by Manus, not only are there automatically generated ranking recommendations, it also categorizes the candidates into different levels based on important dimensions such as work experience. After receiving the user's preference to present the results in an Excel spreadsheet, Manus can also automatically generate the corresponding spreadsheet by writing a Python script on the spot.

Manus can even remember "the user prefers to receive the results in a spreadsheet format" during this process, and will prioritize using the spreadsheet form to present similar task results next time.

Manus can remember the user's preferences in the content generation process | Image source: Geek Park

The second case, more tailored for Chinese users, is the selection of real estate.

In the case, the user wants to buy a property in New York, with requirements for a safe community environment, low crime rate, and high-quality primary and secondary schools - of course, including the most important budget that can be afforded with the fixed monthly income.

In this task, the Manus AI breaks down the complex task into a to-do list, including researching safe communities, identifying quality schools, calculating budgets, searching for properties, etc. It carefully reads articles about the safest communities in New York through web searches to collect relevant information.

Next, Manus writes a Python program to calculate the affordable property budget based on the user's income. Combining the relevant property price information from real estate websites, it filters the property list within the budget range.

Manus can automatically search and filter out properties that do not meet the user's requirements | Image source: Geek Park

Finally, Manus will integrate all the collected information and write a detailed report, including community safety analysis, school quality assessment, budget analysis, recommended property list, and relevant resource links - just like a professional real estate agent. And because Manus has the attribute of "completely based on user's interests", its use and experience are even better.

In the last case, Manus demonstrated its stock price analysis capabilities.

The task given is to analyze the correlation between the stock prices of NVIDIA, Marvell Technology, and TSMC over the past three years: it is well known that these three stocks are closely related, but it is difficult for new users to quickly sort out the causal relationship.

Manus' operation is very similar to that of a real stock broker. It first accesses information websites like Yahoo Finance to obtain historical stock data, and also cross-verifies the accuracy of the data to avoid being misled by a single information source, which could have a significant impact on the final result.

In this case, Manus also used the ability to write Python code, perform data analysis and visualization, and introduced professional financial tools for analysis. Finally, it provided feedback on the causal relationship through data visualization charts and a detailed comprehensive analysis report - just like a "intern" in the financial field doing their daily work.

Not only that, the Manus website also displays more than a dozen scenarios in which Manus can be used: directly using Manus to help you organize your itinerary, personalize your travel route recommendations, and even learn to use various complex tools to streamline your daily work.

In this process, what truly sets Manus apart from traditional tools is its autonomous planning to ensure the ability to complete tasks.

The ability to self-learn also makes Manus's work capabilities more akin to real human beings - even at this stage, it may not be able to achieve expert-level proficiency in a specific field, but its huge potential can already be seen.

With the addition of self-learning capabilities, the versatility of the AI Agent has been greatly enhanced. In user testing of Manus, you can even directly describe the relevant content in a video scene, and Manus can ultimately cross the content limitations of search engines within the platform, accurately find the link to a specific Douyin short video.

Since the current version of Manus is completely based on cloud-side asynchronous operation, the capabilities of Manus are not limited by the form of the terminal platform or computing power you use - users can even turn off their computers after issuing instructions to Manus, and Manus will automatically notify you of the results when the task is completed.

This operational logic is also very familiar - just like an intern who is told "organize the files and send them to me" after work. The difference is that this intern can now respond to you 7x24 hours a day, and you don't have to worry about them "cleaning up the workplace".

02 Multi-agent + self-inspection, running the AI Agent flow

From the above cases, it is not difficult to see that the real trump card of Manus is not the "AI Agent" concept that has already appeared in Computer Use, but its ability to "simulate human work methods".

Compared to "running calculations", Manus's work logic is more like "thinking and executing commands". It has not achieved what humans are currently truly unable to do; this is why some users who have experienced the current version of Manus describe it as "an intern".

On the Manus website, there are many tasks that Manus can complete, including a case that demonstrates how to use Manus in B2B business. Quickly and accurately match your ordering needs with global suppliers.

Manus AI uses an architecture called "Multiple Agent", running in independent virtual machines. Through the division of labor and collaboration mechanism of planning agents, execution agents, and verification agents, it greatly improves the processing efficiency of complex tasks and shortens the response time through parallel computing.

In this architecture, each agent may be based on an independent language model or reinforcement learning model, communicating with each other through APIs or message queues. At the same time, each task is also run in a sandbox to avoid interfering with other tasks, and supports cloud-side scaling. Each independent model can mimic the process of how humans handle tasks, such as first thinking and planning, understanding complex instructions and decomposing them into executable steps, and then calling the appropriate tools.

In other words, through Manus's multi-agent architecture, it is more like a team of assistants, who cooperate to complete tasks such as resource retrieval, connection, and information verification, to help you complete the entire workflow - this is actually not only like you have hired an "intern", but more like you have become a miniaturized "department manager".

In the B2B business case, Manus, through web crawling, code writing, and execution capabilities, automatically searches the vast ocean of the Internet, and matches the most suitable sources for you based on your own needs, in terms of product quality, price, and delivery capabilities. Not only can it present the conclusions to you in a visual way, but it can also provide more detailed operational suggestions for these data.

Manus may be more useful than the built-in tools of a single platform in the B2B scenario | Image source: Geek Park

03 The ultimate "stitching" is explosion

What kind of company is Monica.im behind Manus?

Monica is an All-in-One AI assistant, with product forms gradually expanding from browser plugins to Apps and web pages. The mainstream usage scenario is that when users click on its small icon in the browser, they can directly use the various mainstream models it has access to. By accurately understanding the needs of niche user scenarios, Monica has reaped the "low-hanging fruit" of large models.

Its 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. In 2024, the number of Monica users doubled to 10 million. At the same time, it maintains considerable profitability and is at the forefront of overseas similar products.

Monica's strong performance verifies one thing: Encapsulating to the extreme, it is both TPF and PMF, and ultimately leads to user value.

Monica homepage | Image source: Monica

Manus may have continued the team's approach - Xiao Hong said in an interview that products cannot only have the form of chatbots, and Agent will be a new form that requires new products to support.

He got inspiration from the AI programming products CURSOR and DEVIN. According to Geek Park, the former is mainly in copilot mode, while the latter is in autopilot mode, which is more in line with human needs. The Agent should also be like DEVIN, oriented towards the general public, and truly led by AI to execute. But the problem in the past was that the models were not smart enough.

"I definitely don't know if the Agent can be brought out in this way, it is an unknown thing." he said.

The intriguing thing is that Monica, who "didn't know how to do Agent", has now produced a product that has shaken the entire AI circle.

Manus may not be the ultimate AI Agent, but it has undoubtedly raised people's expectations for AI by another order of magnitude after the explosion of DeeoSeek.

This article is from the WeChat public account "Geek Park" (ID: geekpark), author: shiyun / Zhang Yongyi, editor: Jingyu, published with authorization from 36Kr.