Bottlenecks of AI Agents and the Popularity of AI WorkFlow

Written by
Iris Vance
Updated on:July-13th-2025
Recommendation

The current status of AI Agent technology development and the rise of intelligent workflow.

Core content:
1. AI Agent technology maturity and commercialization challenges
2. AI Agent performance in terms of cost, number of steps and success rate
3. Security issues and the popular trend of intelligent workflow

Yang Fangxian
Founder of 53AI/Most Valuable Expert of Tencent Cloud (TVP)

As we all know, since the introduction of the Large Language Model (LLM), the entire industry has developed rapidly, step by step towards more complex AI agents that can simulate human digital interaction behaviors. However, the commercialization has not been as smooth as expected, and the focus has gradually shifted from AI agents to agentic workflows and data synthesis.

So, why did this change happen? Let's take a look at the reasons behind it.

Agent: The technology is not mature yet, and commercialization is hindered

First of all, many companies have placed high hopes on AI Agents and even invested a lot of resources in research and development. But the reality is very bleak - AI Agent technology has not yet reached the level of large-scale commercial use.

If you often watch some AI Agent demos, you may think they are very cool, but apart from the marketing gimmicks and prototype displays, the actual performance is disappointing. Take the Claude AI Agent Computer Interface (ACI) as an example, its performance is only 14% of that of humans! What does this mean? In short, my guess is better than using it .

The following figure is from TheAgentFactory, which shows the performance of the current AI Agent in terms of cost, number of steps, and success rate. As you can see, the success rate of the AI ​​Agent is only about 20%, while the cost and complexity are very high.

The Operator recently launched by OpenAI has an accuracy rate of 30%-50% in computer operation and browser use, but this is still lower than the human level of more than 70%.

In addition, there is a potential security issue: AI Agent is vulnerable to malicious pop-up attacks when browsing the web, which undoubtedly increases the risk in practical applications.

Currently, there are two main ways for AI Agents to implement tasks: one is through a browser (such as Webvoyager, OpenAI Operator, etc.), and the other is directly through the graphical interface of the operating system (such as Anthropic). Both methods essentially use the GUI as an API. In the early days, attempts were made to develop an API for each application separately, but this approach was not realistic due to the high development cost and the fact that many commercial applications did not have ready-made API support.

Focus on Agentic Workflow

Of course, the problem of AI Agent does not mean that the entire industry is stagnant. On the contrary, more and more companies are beginning to focus on another direction - Agentic Workflow.

We all know that knowledge workers are currently inefficient. A report shows that people spend an average of 30% of their time looking for information. In addition, when faced with complex problems, they need to extract and integrate information from multiple documents, which undoubtedly increases the difficulty of the work.

Agentic Workflow was created to solve these problems. It breaks down complex tasks into simple subtasks and connects these subtasks in series to form a process. As shown in the following figure:

The advantage of this model is that it not only improves efficiency, but also brings explainability and inspectability. In other words, users can clearly see how each step is completed, making it easier to understand the results.

At the same time, with the explosive growth of data, how to effectively integrate and utilize this data has become a key challenge. Agentic Workflow performs well in this regard, and it can help knowledge workers quickly synthesize the required information and generate answers for specific scenarios.

For example, ChatGPT’s Deep Research feature is a great example of this. It can complete multi-step research in a short period of time, solving tasks that might take a human hours to complete.

Here we have to mention a new concept - Agentic RAG (proposed by LlamaIndex). Its core idea is to provide customized data synthesis services for "single audience" at a certain point in time. In the next few months, personalized workflows, information synthesis and desktop orchestration will become hot areas.

Improvement of reasoning and problem-solving skills

Modern AI models are gradually adopting reasoning as one of their core functions, enabling them to systematically handle each link by breaking down complex problems into small parts.

This approach not only improves the efficiency of problem solving, but also enhances transparency, making it easier for users to understand how conclusions were reached.

In the past, users needed to manually add reasoning logic to the prompt words to guide the model on how to break down complex or compound tasks; but now, AI has a certain degree of autonomous reasoning capabilities, which greatly reduces the threshold for use.

Summarize

Finally, I want to emphasize that whether it is RAG, Prompt Engineering or other tools, companies should avoid blindly chasing trends, but return to the essence of solving real business problems. After all, the value of technology does not lie in how advanced it is, but in whether it can create practical benefits.

Whether it’s improving customer experience, optimizing operational processes, or responding to social needs, we should ask ourselves a question: How can we use technology to provide meaningful solutions?

Only in this way can enterprises remain competitive in the future and not be eliminated by the rapidly changing technological wave.