Has an AI agent that can work like a team of human experts emerged?

In 2025, the buzz around the "Year of AI Agents" shows no sign of fading, with star products emerging frequently, making it the most imaginative track in the current artificial intelligence field.

But after more than half a year, has Agent really become more user - friendly?

On one hand, there is the dilemma of Manus. On the other hand, the experience of a group of Agent products born with a "silver spoon" fails to meet expectations. The uneven performance in tasks forces users to switch between technological show - off and "manual intervention", wasting a great deal of optimization time.

AI Agents can generally handle tasks at a single point, but they are still far from "doing a good job". As the tasks people assign to AI become more and more complex, what is needed is no longer just linear delivery, but a "system - level" player that can break through in specific scenarios.

01

Complete over 5 complex tasks simultaneously in just 3 minutes

Currently, the core bottleneck faced by AI Agents is neither computing power nor cost, but the single - threaded serial architecture.

Single - threading means that the invocation of each task and each AI is strung together with linear thinking. All requests must be processed strictly in sequence. This architecture is inherently unable to think about several complex problems simultaneously like a human being, dynamically adjust task priorities, and think and execute tasks concurrently.

This linear thinking makes it difficult for Agents to understand users' complex needs, and it is also very challenging for users to describe them.

In addition, the processing speed of single - threading is extremely slow. Since it operates unidirectionally along a blocking chain of generation, waiting, and the next generation, any blockage in any link will cause a global jam. In terms of architecture, this unidirectional Agent, like a factory conveyor belt, can never meet users' requirements for efficiency, experience, and delivery quality simultaneously.

Not to mention that in terms of context memory ability, many Agents have not yet evolved from tools to knowledge bases, and they simply cannot achieve personalized and precise matching between task execution and delivery standards.

These three pain points are already quite "fatal". Unfortunately, the AI Agent field is also suffering from high traffic costs, and the queuing for invitation tests wears down users' patience. This is why, after Manus became popular, large companies launched their own Agent products at the fastest speed, hoping to be the one to break the deadlock.

In the past few months, during the exploration and iteration of various parties, the performance of Wenku GenFlow 2.0 has been particularly outstanding.

First of all, as a general - level Agent with greater imagination, Wenku GenFlow 2.0 has achieved "full - end availability": it is one of the earliest full - end general Agents in the world. It has been fully launched on the Web and App of Baidu Wenku, without the need to queue with an invitation code, and it is free for a limited time.

Behind this, based on the insight into users' real workflows, Baidu Wenku's network disk has self - developed an innovative Multi - Agent architecture. It abandons the logic of an intelligent agent where a "super brain" takes care of everything, and instead builds an "AI expert group" composed of more than 100 vertical - field expert Agents, enabling them to complete tasks in a parallel and collaborative manner.

Wang Ying, the person in charge of the Wenku Division and the Network Disk Division

Relying on the Multi - Agent and the underlying MoE (Mixture of Experts) technology, GenFlow 2.0, which is not restricted by "congestion", not only delivers surprising results but also improves both the quality and efficiency of processing complex tasks.

This breakthrough and evolution in architecture have completely expanded the ability boundary of "AI working like a smart person". For example, it has upgraded from taking dozens of minutes to generate a PPT or a document to completing over 5 complex tasks in parallel in 3 minutes and delivering a cross - modal solution, making the imagination of AI application implementation no longer a "winged horse".

As the general ability improves, the barriers in different scenarios are gradually being weakened. Full - end universality represents a richer software and hardware ecosystem and "boundless" cross - end collaboration. Especially on the mobile side, which is often neglected in AI office applications, users can modify plans on the subway, extending work from the desktop to fragmented scenarios, so that users will no longer feel that AI work is a waste of time.

The Agent workflow of "inputting requirements and getting solutions" is becoming simpler and more convenient.

02

Refresh the form of AI human - machine interaction: No longer "finding tools", but "assembling an expert team"

The "way to break the deadlock" of Wenku GenFlow 2.0 is not limited to the improvement of architecture, delivery quality, and speed. More importantly, the underlying logic of human - agent collaboration has even gone beyond the narrow concept of "Agent".

Agent is translated as an assistant, aiming to complete the instructions of the issuer. The reason why Wenku GenFlow 2.0 focuses on the Flow workflow is that it has identified a blind spot: in reality, most "human assistants" are not business experts, but are good at integrating and distributing for decision - makers, assigning different tasks to different professional teams.

However, completing these tasks definitely requires front - line business capabilities. Therefore, two indispensable points in the concept of Wenku GenFlow 2.0 are:

First, the precipitation of public and private domain data and user memory libraries, and the continuous accumulation of user profiles and preferences;

Second, a professional AI Team.

In this way, GenFlow 2.0 becomes a scheduling center, autonomously planning and dynamically scheduling the expert Agent group according to requirements. With just a few words from the user, a professional AI team of a hundred people that "continues to evolve" can be mobilized.

Take industrial design ready for production as an example. When a user inputs "Design a set of Crayon Shin - chan blind boxes for me", how does Wenku GenFlow 2.0 "assemble a team"?

First is requirement analysis. GenFlow 2.0's intelligent understanding and mode switching can accurately identify the user's intention and autonomously switch the collaboration mode.

The system will first identify the requirement of "creating 5 - 8 3D Q - version character designs based on the IP image of Crayon Shin - chan", then autonomously think and plan the path, dispatch appropriate design Agents to generate sketches, infer the style and scenario based on the user's preferences presented in Baidu Wenku's network disk and active inquiries to the user, and conduct production cost accounting simultaneously. This step is more like an Agent assembling a project team, which is supported by the intention understanding and multi - task parallel ability of the Multi - Agent architecture.

During the execution process, Wenku GenFlow 2.0 allows direct intervention throughout the task. Users can pause at any time, add new requirements, and call files from the network disk to optimize and control the generation process and results in real - time. In terms of interaction, it puts users first. For example, when the system understands the intention as "different images of Shin - chan", if the user inputs "Introduce other characters from Crayon Shin - chan", telling the Agent that there are more characters than just Shin - chan in this set of blind boxes, GenFlow 2.0 will automatically search for other characters from Crayon Shin - chan across the network and select appropriate images.

At this time, if you want to create a product plan for the generated blind - box pictures, you just need to add the requirement for creating a PPT, and you can get both the pictures and the PPT output. The PPT can also be regenerated according to the outline and edited in real - time.

The ability to intervene throughout the task process subversively solves the pain points of the "black - box" generation process and uncontrollable results, and also provides the knowledge base with the value of being browsed at any time like a reference document. The memory library and personalized content are dedicated to continuation and high - quality delivery. Files in the network disk can be called, and relying on the self - developed AI editor, full - process editing can be carried out anytime and anywhere, realizing a closed - loop from the starting point to the end point of creation.

Why is it Baidu Wenku's network disk that has an AI expert group first? The answer lies in the two - year AI reconstruction. Most of the leading Agents in Baidu Wenku's network disk are self - developed, and all multi - modal Agents have been verified by hundreds of millions of users of Baidu Wenku's network disk. In addition to the intelligent PPT Agent with the world's leading access volume, Wenku GenFlow 2.0 has reached the expert level in many mature leading Agents. For example, in scenarios such as generating long research reports with tens of thousands of words and professional visual charts, text - to - video picture books, text - to - posters, and in - depth and academic searches, it shows the characteristics of multi - modality, high - quality, and deep search.

The amazing performance of Wenku GenFlow 2.0 is closely related to Baidu Wenku's network disk's adherence to the MoE architecture from the very beginning.

The core breakthrough of MoE is that it decouples the relationship between parameter scale, efficiency, and computing cost. Since only a small number of experts are activated for each task while others remain "dormant", the cost - benefit ratio is very high. From the perspective of the same - level density model, the inference cost - performance ratio of the MoE architecture may be several times that of other models.

The effectiveness of the MoE mechanism for Wenku GenFlow 2.0 is not accidental. It is an inevitable technological choice to support Multi - Agent to achieve high efficiency, low cost, and high scalability simultaneously. It also solves the bottleneck of the general model of being "good at everything but mediocre at all" through vertical division of labor, helping Baidu Wenku and its network disk to replicate benchmark cases in various industries.

In terms of scalability, GenFlow 2.0 also provides "Lego - style" capabilities for Multi - Agent, enabling it to access third - party Agents through standardized protocols such as MCP, or connect the capabilities of Baidu Wenku's network disk to the ecosystem in the form of MCP Server. This is a very important step for AI Agents. After all, human experts also become experts through continuous communication and connection.

Currently, Wenku GenFlow 2.0 can not only access Baidu's ecosystem but also flexibly access third - party service ecosystems through MCP compatibility. It is no longer a fantasy of "strongly relying on a single AGI", but a sustainable productivity that combines various resources.

03

The national - level application is moving towards an open ecosystem

The experience upgrade of GenFlow 2.0 is not "groundless". Its core foundation is the next - generation content - field operating system, Cangzhou OS, launched by Baidu Wenku's network disk in April this year.

This system is divided into a three - layer architecture: the underlying infrastructure, the central system, and the application service. It has become the core system for Agent scheduling, Agent ability output, public and private domain data management and processing, and ecosystem connection, reconstructing the way and value of human - machine collaboration:

The underlying infrastructure layer processes public and private domain content by constructing tool frameworks and knowledge - based frameworks, and realizes file parsing, transcoding, searching, and cross - modal content understanding. The central system layer, based on the GenFlow 2.0 scheduling center, combines user memory and portrait data to efficiently allocate and schedule multiple Agents for parallel collaboration. The application service layer integrates hundreds of Agents from Baidu Wenku's network disk and third - party Agents to form a task - closed loop.

Cangzhou OS enables GenFlow 2.0 not only to serve users well but also to generate great value in the B - end market. Through the forms of MCP Server and Agent to Agent, it fully opens up the capabilities of Wenku and the network disk, supporting manufacturers, enterprise users, intelligent agent applications, developers, etc. to access at low cost and high efficiency. For example, after Samsung mobile phones access the MCP Server, they can directly call the file upload and content understanding functions of Baidu Network Disk, solving the difficulty of processing large files on mobile devices.

Manufacturers such as Honor have gone even further. They have natively integrated GenFlow 2.0 into Honor's intelligent assistant YOYO through MCP, realizing system - level native scheduling between AI Agents and hardware manufacturers. Through the MCP ecosystem and the intelligent agent scheduling ability of GenFlow 2.0, users of Honor MagicOS can obtain their personal network disk knowledge base and professional documents from Wenku with one click, enjoying high - quality experiences such as network disk retrieval, content sharing, online search, picture understanding, file summarization and Q&A, and Wenku PPT generation.

After accessing GenFlow 2.0, Honor has become one of the world's first hardware manufacturers to access the MCP ecosystem, taking the lead in realizing the synergy between AI native intelligence and hardware native capabilities. The standardized ecosystem output by Cangzhou OS has also been extended in terms of scenarios and customers with the recognition of leading hardware manufacturers.

The essence of Cangzhou OS is neither to strictly control hardware resources like traditional operating systems nor to have obvious limitations in single - point capabilities like single - Agent products on the market. It realizes standardized connections between Agents and between Agents and external services through MCP, just like creating a "universal language" for the AI content world, enabling GenFlow 2.0 to derive an intelligent agent scheduling mode of an "expert group" on this basis.

In this way, any enterprise's AI requirements can feel the flexible adaptability that can fit into various "containers" like water. This breadth and flexibility are driving Baidu Wenku's network disk from a