Image source: Generated by Unbounded AI

The AI ​​circle is ushering in a new round of technological explosion period. DeepSeek shocks the world and abroad with its deep reasoning model R1. OpenAI, Anthropic, Google, Alibaba, Byte, Baidu, etc. have promoted new developments in their own big models. In addition to the model layer, the application layer has also begun to have some new changes.

Recent innovations come from a new product launched by a founding team of Chinese post-90s, Manus, an application tool known as the universal AI Agent.

Manus' official website (https://Manus.im/usecases) gives some examples for us to understand Manus's capabilities more intuitively. It can be seen that its functions are indeed particularly powerful in the official demonstration. Currently, the use of Manus requires the invitation code to be obtained in advance.

It is necessary to briefly introduce Manus. Compared with traditional AI assistants, similar to Doubao, Kimi, and DeepSeek are limited to information retrieval and suggestions interaction, while Manus has realized the closed-loop capability of "self-execution of complex tasks → delivery of complete results" for the first time, filling the current market gap.

One other point is the breakthrough in the definition of universality: dealing with complex tasks across fields (such as stock analysis, travel planning, code generation, etc.), breaking the limitations of vertical track AI‌, which is the first two amazing points. The third point is that Manus is another masterpiece after the Chinese team made DeepSeek. It is positioned as a general assistant for "using both hands and brains". It emphasizes the unintermediated execution model of "input requirements → output results". It is named from the Latin word "Mens et Manus" (meaning both hands and brains). The technical concept advocates "Less structure, more intelligence", reduces the dependence of preset processes and strengthens the independent evolution ability of the model.

01 , Advantages and Disadvantages of Manus

Summary of the core advantages of Manus:

Independent execution of the entire process: from demand analysis (such as resume screening), task dismantling (such as real estate purchase planning) to results delivery (generation of Excel/PPT reports) no manual intervention is required.

Multi-tool chain integration: supports browser operations, programming tool calls, cross-platform data crawling, and realizes complex operations (such as automatically decompressing resume packages and analyzing).

Covers many fields such as education (teaching material generation), finance (stock relationship analysis), human resources (candidate grading), enterprise services (automation weekly report).

Cross-domain collaboration: two heterogeneous tasks such as resume screening and real estate purchase can be handled simultaneously, reflecting the characteristics of general Agent‌.

Multi-Agent Collaborative System: Planning, Execution and Verification are separated, and cloud virtual machines handle tasks in parallel.Improve fault tolerance.

The above are the advantages of Manus, and Manus also has some disadvantages:

Rely on the preset process framework: Some tasks need to be executed in the "restricted environment" between the browser and the operating system, and cannot be adapted to unpredefined applications.

Execution stability problem: User feedback has uncontrollable behaviors such as generating false results (such as forging invitation codes) before crash.

The high hype popularity has led to a gap between early user expectations and actual capabilities, and some scenarios still require manual correction of output results‌.

Enterprise deployment cost: The cloud virtual machine operation mode has a high demand for computing power, which limits the use of small and medium-sized enterprises.

The legal effect of self-signing a contract has not yet been clarified, and cross-platform calls to privacy data may cause compliance disputes.

Workplace replacement anxiety: 80% of mid- and low-frequency white-collar workers (such as resume screening and report generation) face the risk of being replaced by low-cost AI.

02. What technological innovations has Manus made?

After talking about the advantages and disadvantages, let's focus on what the technological innovation behind Manus's strength is.

1. Multi-Agent Collaboration System (Multi-Agent Collaboration)

Three-layer architecture design: Planning Agent: Use Monte Carlo Tree Search (MCTS) algorithm to dynamically disassemble task priorities, and support real-time adjustment of execution paths (for example, priority treatment of academic qualification matching in resume filtering).

Execution Agent: Integrates 200+ tool interfaces (browser, Office suite, etc.), simulates human operations (clicks, scrolls, and form filling) through the browser automation engine, but is limited to CS architecture applications that do not open APIs.

Validation Agent: Deploy adversarial testing modules to detect logical contradictions (such as triggering the review process when the financial data deviates from the industry benchmark by more than 5%).

Multisig System: Multisig System: Multisig System is used to process subtasks in a collaborative manner to improve the reliability of the results through cross-verification (for example, in stock analysis, 3 models independently calculate the industry correlation and obtain the highest confidence value).

2. Dynamic training and optimization mechanism

Real-time feedback iteration: Users can use the "dynamic training" function to correct the output format (such as specifying PPT templates) or adjust the execution path (such as skipping data crawling of specific websites), and the model is updated instantly.

Crash recovery technology: Automatically generate alternative solutions when the task is interrupted (such as changing to keyword matching mode after the resume parsing fails), and keep the progress to the local cache.

3. Deep integration of toolchain

BrowserSandbox technology: Run cross-platform operations in an isolated environment (such as automatically logging into the recruitment website to crawl resumes) to avoid user privacy data leakage.

Memory preference system: Continuously learn user habits (such as contract template preferences), occupy cloud storage resources to form personalized execution strategies.

4. Performance verification system

GAIA benchmark test (https://zhuanlan.zhihu.com/p/669652697) breakthrough: in the third-level tests, reaching 86.5%/70.1%/57.7%, respectively, surpassing OpenAI's DeepResearch model and approaching the human level (90%).

Real scene verification: implement data analysis automation in the Upwork platform to complete design tasks (such as Logo generation) and Kaggle competitions to verify cross-domain execution capabilities.

If the above Manus's technological innovation makes you shine, do you have a question at this time, why is it Manus? Why did other AI manufacturers, especially large technology manufacturers, not achieve such results? Is other manufacturers not able to have such capabilities? At present, the reasons may be as follows: 1. ‌Technical Path Reliance‌

‌Model Scale Priority Strategy‌: Manufacturers such as OpenAI focus on expanding the amount of model parameters (such as pursuing GPT-5), and have not regarded tool calling capabilities as the core R&D direction.

‌Rules Engineering Inertia‌: Traditional AI products rely on preset processes (such as customer service dialogue trees), and it is difficult to adapt to the "Less structure, more intelligence" concept advocated by Manus.

2. Engineering Challenges

‌Multi-tool interface adaptation cost‌: It is necessary to develop special drivers for browsers, office software, etc., large companies are unwilling to do dirty work, and other small manufacturers lack ecological integration capabilities (for example, the Manus team took 2 years to complete the development of 200+ tool interfaces).

‌Long-period task management problem‌: Traditional models are limited by conversational interaction, making it difficult to realize the asynchronous working mechanism of “users continue to execute cross-daily tasks after offline”.

3. Market positioning differences

‌B-side service barriers‌: Microsoft Copilot and other products focus on Office scenarios and lack cross-field generalization capabilities (such as handling resume screening and stock analysis at the same time).

‌Ethical risk avoidance‌: Leading manufacturers are afraid of legal disputes caused by independent decision-making (such as the effectiveness of contract signing) and choose a conservative technical route‌.

03 , Manus's subsequent improvements and industry impact

In view of some shortcomings currently shown by Manus, we believe that its future improvement points are as follows:

Enhanced environmental adaptability: break through the "mezzanine" limitations of the browser/operating system and implement lower-level system permission calls.

Reduce computing power dependence: promote the popularization of small and medium-sized enterprises through the combination of model compression and edge computing.

Prioritative implementation scenarios: enterprise automated office (intelligent customer service, contract review), personalized education teaching (dynamic generation courseware).

Open source ecosystem construction: open some model interfaces to attract developers to build plug-in ecosystems (such as third-party tool integration).

Establish an AI execution audit mechanism: leave traces on the independent decision-making process and clarify the responsibility for wrong results.

Data security specifications: Restrict cross-platform circulation permissions for sensitive information (such as resumes, financial statements).

But it cannot be ignored that the emergence of Manus will definitely have some impact on other manufacturers: forcing OpenAI and other manufacturers to accelerate the pace of shifting from "dialogue interaction" to "task execution" product development, and accelerate the iteration of Agent technology; promoting multimodal interaction (text/chart/code mixed output), and tool calling API integration have become the new benchmark for AI products.

According to the Manus plan, the subsequent official will also have an open source reasoning module, which is expected to spawn a secondary development community based on its technology stack‌.

The surge in demand for Manus may also drive a surge in demand for edge computing devices, causing AI terminal deployment to extend from the cloud to the local area, and a group of domestic hardware manufacturers are also expected to benefit from it.

The speed of development of AI is driving a reshuffle of all walks of life. The future is not an era of assets, nor is it an era of power, but the future will be an era of technology.