Manus would be according to benchmarks and the feedback from first testers, the most independent IA agent currently on the market.

New advance in agentic AI. The Chinese company Manus, owned by the Singaporean Butterfly Effect group, unveiled on March 5 a completely autonomous AI agent. In a 4-minute video, Yichao ‘Peak’ Ji, co-founder and chief scientist, presents the main use cases: independent information search, financial analysis, prospect identification … The agent, available in early access, is based on a very little documented technical stack. But the results are there.

Manus, a truly autonomous AI agent

Manus presents itself as an entirely autonomous agent, capable of performing complete tasks without constant human intervention. The agent works asynchronously in the cloud. He continues to work even after closing the site. The AI ​​can in particular analyze documents, process files (such as CVs), create detailed reports, develop spreadsheets, and even develop interactive websites from analyzed data.

Manus can interact with different digital environments. The agent can browse the web (with a dedicated chrome version) to search for real -time information. He is also able to write and execute Python code for data analyzes, deploy websites, and even, according to his designers, to participate in KAGGLE competitions.

For complex tasks, manus begins by decomposing the project into smaller steps, creating a list of tasks which he then performs in a methodical way, like an experienced human assistant. Unlike Operator of Openai, designed to collaborate more closely with the user, Manus seems to make decisions alone. The interface is very simple to use: the user gives a quick to chatbot And the AI ​​performs the necessary actions step by step.

In a demonstration presented by its designers, Manus processed a file containing 10 CV, then 5 additional. The agent automatically decompressed the files, traveled each CV page by page, and saved the important information in structured documents. After analyzing all 15 applications, manus provided a classification of candidates with their detailed profiles and the evaluation criteria used, before transforming this data into a spreadsheet at the user’s request.

How does Manus work?

Manus works as a multi-agent system, powered by several distinct models operating in coordination. The agent begins by analyzing the user’s request, then creates a detailed execution plan in the form of a task list. For each step, Manus uses specific tools: web navigation, code execution, data analysis or content creation. The precise technical infrastructure of manus remains little documented, but according to the first returns of several users, the agent would use a combination of Claude 3.7 and other models of the Qwen Fine-Tués family. A pragmatic approach that would allow him to offer advanced capacities without the teams having had to fully develop a new foundation model.

Another novelty, Manus uses an Ubuntu terminal to execute certain tools under Linux, a first in the field. The agent can thus run Shell commands, install packages, manipulate files like a human would. By combining access to a web browser, Python programming environments and a Linux terminal, manus manages to make complex workflows which would normally require several separate software and considerable technical expertise. The tool goes much further than most agents already deployed on the market.

3 stunning use cases

The first user feedback with access to the closed beta version of manus show the extent of the possibilities. From financial analysis to programming, shared demonstrations on social networks illustrate the versatility and efficiency of AI. A user reports having asked Manus to carry out “a professional analysis of the Tesla action”, a try that the agent would have accomplished in about an hour, producing work equivalent to “about two weeks of professional work”.

In the field of programming, another testimony describes manus as “the most impressive AI tool ever tried”. The User (the Hugging Face Product Manager) simply asked the agent to “code a Three.js game where an airplane is controlled”, and Manus has performed the development task without requiring additional interventions.

Finally, the third example illustrates the creative capacities of Manus, which managed to “search for video games recently coded by AI and create a website about them in the style of a video game from the 90s”.

The already tangible impact of manus

Despite a technical infrastructure based on models developed by other companies, Manus generated a considerable impact within 48 hours of its presentation. Influential figures in the sector quickly recognized its potential. Ethan Mollick (professor at Wharton and worldwide recognized in AI), Jack Dorsey (founder of Twitter) or Victor Mustar (Product Manager at Hugging Face) have publicly stated on X that they were amazed. According to our colleagues from Bloomberg, the announcement of Manus AI on March 6 even led to an increase of 7% of Alibaba’s shares.

Not to mention new time Deepseek, manus marks a significant advance in agency AI by skillfully combining existing technologies. No commercial offer has yet been revealed, the access remaining limited to a closed beta, but a rapid deployment could seriously shake up the market that coveted Openai and other American AI giants.