The Chinese AI agent Manus brings web search, Ubuntu sandbox and VS Code

Contents

Chinese manufacturer Manus is marketing its new AI agent as “potentially a glimpse at AGI”. AGI, Artificial General Intelligence, is the Holy Grail of AI research, as it is supposed to have human-like intellectual abilities and thus be able to perform any task.

Beyond the marketing promises and a series of application examples on the manufacturer's website, little is known about the inner workings of Manus. In the promotional video on the website, Manus calls itself the “General AI Agent”, which is a “multi-agent system based on several models”. The video also announces that some of these models will be released as open source at the end of the year. Especially those “post-trained for Manus”. These are unspecified GenAI models that the agent uses to solve tasks. However, it is not known which models are behind them, whether the manufacturer generated them himself or how he trained them.

To test Manus' capabilities, we gave the agent a series of tasks, which it – even if we cannot speak of AGI – completed with quite remarkable results and showed some special features. For example, the AI uses an Ubuntu sandbox in which it tests code and even installs dependencies on its own. Manus presents coding results in a cloud version of VS Code. And the agent also solved a classic GenAI task of creating a comprehensive strategy for digital sovereignty quite acceptably – albeit with the usual limitations that apply to GenAI in general: Manus fantasizes a Covid reference into the concept that is no longer valid in 2025.

Manus, nosce te ipsum

First, we asked Manus the question about himself:

Manus can be considered a multi-agent system from an […] architectural perspective, as it internally uses different specialized modules […] that work together:

1. a planner module that analyzes tasks and plans steps
2. a knowledge module that provides relevant information and best practices
3. a data source module that provides access to various sources of information

Manus emphasizes:

“Manus-AI is characterized by its ability to handle a wide range of tasks – from information research and fact checking to data analysis, complex content creation and programming. […] From the user's perspective, Manus appears and behaves as a unified agent with a coherent personality and interaction model.”

According to the promotional video, Manus can create and execute small, problem-related Python programs, for example to calculate the budget for a move. It can also independently analyze online resources. Manus uses an Ubuntu sandbox for related work, presumably in a virtual environment or container.

In a chat reply, Manus tells us: “I checked the system and found out that the sandboxes work with Ubuntu 22.04 LTS (Jammy) instead of Ubuntu 24.10 as mentioned in your request.”

In several attempts: Overlaying triangles with CUDA

For the first programming task, Manus receives a prompt from the comparison of DeepSeek R1 and ChatGPT o3-mini-high.

The AI is asked to design a program that repeatedly overlays a predefined set of semi-transparent triangles on a GPU using the CUDA programming interface developed by Nvidia to create a candidate image. For each color channel of each pixel, it should then calculate the deviation from the corresponding pixel of a target image and finally output the candidate image with the lowest total deviation in PNG format.

However, the program initially presented by Manus could only be translated after a few modifications. The supplied CMakeLists.txt file had to be adapted to use the nvcc for the translation of “CUDA-containing” code (CMake can trigger this automatically with suitable file extensions “.cu”). This code was saved in a file with the extension .cpp, so it had to be renamed. The resulting image had the correct dimensions, but the triangles seemed to be too large and only come from one color spectrum. A second attempt with an identical prompt delivered significantly more complex code – different solution approaches with multiple calls are normal for AIs – which could not be translated despite the correct .cu endings.

Since a quota was already noticeable here after a few initial attempts – Manus is obviously currently subject to a very high level of interest – the example was postponed for now. Such problems are to be expected with a beta. Further work with Manus was more successful, and the triangular example also appears again at the end.

Visual Studio Code on the web

A special feature of Manus was already noticeable during the first attempts: You have optional access to an instance of Visual Studio Code running as a web interface while you are working. Although this allows direct insight into the generated code (Figure 1), it does not allow much interaction. All work appears to be carried out in an Ubuntu sandbox. Manus also loads code here if required.

Such a sandbox corresponds to a “task” in Manus terminology. Users can also submit several queries in succession or refine the results, as is the case with DeepSeek, ChatGPT and others. The associated AI tasks are each assigned to a sandbox.

This is practical because the beta quota seems to be implemented via the number of permitted tasks. However, after some more complex questions, the system sometimes indicates that the context size has been exceeded, which may be freely interpreted as exceeding the permitted resource usage. In this case, however, Manus was able to access the data of a stopped sandbox assigned to the account from another sandbox and process it further, at the cost of consuming a new task.

In another case, the generation stopped in the middle of a promising attempt. For example, the main.cpp then contained the text

<response clipped><NOTE>To save on context only part of this file has been shown to you. [...]</NOTE>.

This was frustrating and probably due to the overload of the beta infrastructure. As with most other AIs, interaction can take place in either English or German (and presumably Chinese), although English and German responses were always grammatically perfect.

A brief attempt to have Manus develop a UML class diagram for one of the author's C++ classes using PlantUML was successful. The diagram showed the correct information.