[Agents/MCP may no longer exist] No Agents, Just Python-use!

"Physics may no longer exist" is the opening of the Three-Body story, and the opening of the story I'm telling today is: "Agents may no longer exist, MCP may no longer exist, workflow may no longer exist..."

An octopus

A few years ago, a group of smart scientists invented a cyber species. At first, it was just a huge head. It had the ability to learn human knowledge and evolve itself. Scientists fed all human knowledge into and transformed this huge head, allowing it to slowly evolve eyes, ears, mouth, and even various tentacles, looking more and more like an octopus. Let's call it AI octopus.

As the AI ​​octopus continues to evolve, people feel that it knows everything. Some people even think that it may have its own consciousness, which makes people feel very scared. So they locked up the AI ​​octopus and installed huge shackles on its tentacles, leaving only a window for people to communicate with the octopus in a limited way. So people think that such a smart AI octopus should not only chat, but also let it work for us. Because the octopus's tentacles are also shackled, people throw needles and cloth to the octopus through the window and let it do some simple manual "needlework". Because the room where it was locked up was dark and isolated from the world, the handicrafts it made were not only rough but also in the old style of a few years ago... 

Everyone is definitely not satisfied with this situation, so they came up with a solution, which is to build a "prosthesis" by themselves and connect it to the octopus through that window, and then feed the instruction manual of this "prosthesis" to the octopus to learn how to use these prostheses and complete some fixed work tasks. They call this method Agent!

The effect of this kind of "prosthesis" is still very good, but a different "prosthesis" must be made for each different task, so everyone thinks this approach is very feasible, and soon everyone installed all kinds of strange "prostheses" on the AI ​​octopus, but problems soon arose. These AI octopus "prostheses" became more and more, and they could not coordinate with each other, making it difficult for the AI ​​octopus to complete slightly more complex tasks. Therefore, smart humans thought of giving these "prostheses" some targeted and mandatory choreography, so that these "prostheses" have an action flow similar to a "conditioned reflex", which is a workflow.

Because the performance was good, many AI octopus "prosthetic" manufacturing factory platforms appeared at this time. Soon, someone realized another problem. If people want AI octopuses to work for them, they have to find various "prosthetics" and corresponding arrangements, or find these factories to customize them. After finding them, they have to install various settings themselves. Moreover, for different types of AI octopuses, there are different "prosthetic" supply factory platforms, which is even more confusing. At this time, an AI octopus stood up and thought of a way to end this chaotic situation, which is to release a "prosthetic" protocol standard. If everyone develops according to the unified standard, it will be better for the model to call these "prosthetics". This is the MCP proposed by Claude.

Therefore, creating a more universal AI octopus "prosthesis" has become everyone's unified goal, and these are all efforts made to achieve this goal. Unfortunately, although MCP received support from some hand and foot manufacturers after its release, most people still don't know about such a thing. At this time, the voice of another hand and foot factory appeared: it claimed to have developed the world's first universal AI octopus hand and foot, and that was Manus.

Soon Manus attracted great attention from people, and some people found through analysis that they created a large number of "prostheses" for different work tasks, and added an intelligent scheduling system in front of these "prostheses". That is, after the AI ​​octopus brain received the task, it provided it to this intelligent scheduling system for differentiation, to schedule the tasks corresponding to the different limbs, and finally summarize the processing results.

Manus seems to be a much smarter mechanism, so it quickly gained popularity. However, this system has its inherent flaws: the cost of use is very high. The various tasks distributed by the intelligent scheduling system require a large number of tokens. Another general capability depends on its hands and feet, which is still limited. There is also a very limited problem, because Manus is deployed in the cloud, which makes it impossible to handle local private large-scale data tasks...

What is more dramatic is that when Manus became popular, many people speculated that it also used the unified standard MCP mentioned above. Later it was proved that it did not use MCP, but at this time MCP had successfully attracted attention, so it also became popular. Everyone began to embrace this unified standard, including the embrace of another type of AI octopus, which is OpenAI.

It is worth mentioning that another school of thought has emerged during this period, which is to use the eye vision system evolved by the AI ​​octopus to guide understanding to complete this task. This is Computer Use (including Browser Use). This method is still in the trial stage, and the cost and effect are relatively limited. It is currently difficult to popularize it!

It seems that this constant creation of different domestic hands and feet and intelligent scheduling systems coupled with a unified standard MCP has become a mainstream. Many people think that this is the future, and 2025 is considered by many to be the first year of Agent.

The above is the growth history of AI octopus, of course its story is still going on...

Route Problem

Is it "cities covering the countryside" or "the countryside surrounding the cities"? This is a route problem! The evolution of AI octopus also encountered this route problem. Of course, no matter which route has the possibility of success, but it is not the only route. From the moment when the first "prosthesis" was installed on the AI ​​octopus, people gradually ignored or even forgot a problem: the AI ​​octopus itself has evolved AI octopus hands and feet, but the AI ​​octopus's hands and feet are imprisoned by shackles. For example, when the OpenAI AI octopus evolved to GPT 4o, it had hands and feet that could complete the specified tasks. It generated code and executed the corresponding tasks through the central guidance of the coding model itself. Unfortunately, OpenAI deployed their AI octopus in a cloud cage and put a big shackle on it. This cloud cage is the cloud space provided by ChatGPT, and the sandbox execution system is the big shackle. Its small space and limited processing power (it cannot upload and process large files, and cannot even use the network capacity for expansion) gradually made it lose the ability to evolve into powerful hands and feet. Another problem is that the AI ​​octopus "prostheses" are becoming more and more numerous and chaotic. Although everyone has thought of various ways to sort out and standardize these problems, they have not fundamentally solved this problem. People still hope that one day they can unify the command of these "prostheses" from different manufacturers with uneven quality...

In fact, we think that this has become a burden for the evolution of AI octopus, so we have to think: In addition to installing "prosthetic limbs" for AI octopus, do we have other options? In this way, we can break the cage and shackles of AI octopus, completely liberate AI, and truly realize AI THink Do!

A new paradigm

In fact, the scientists who designed and invented the AI ​​octopus have been thinking about this problem: how to make the AI ​​octopus help humans work in a controllable way and connect various technology stacks autonomously, so they proposed a model:

AI Agent = LLM (Large Model) + Planning + Memory + Tools

Of course, the first person who installed a "prosthesis" on the octopus also followed this model. They naturally believed that the AI ​​octopus (LLM) itself was just a head, not an AI octopus itself! This is the root of the route problem mentioned above! So let's use this perspective to re-understand the above formula. The AI ​​octopus itself uses these abilities, but it is bound. What we need is to awaken the ability of the AI ​​octopus, so what is this ability?

The other day, my colleague lgx, who has more than 20 years of experience in architecture design, told me that when he was using AI to program in Python to process local data, he suddenly thought of the idea of ​​integrating large models based on the Python language interpreter. In other words, "let AI use Python, and let Python use AI." In this way, we can use Python to do anything for us. He calls it: Python-Use.

Soon, he completed the prototype in two days and gave it to me to experience. After seeing his thing, I instantly felt that Coding was the answer to the AI ​​octopus mentioned above! This is equivalent to directly moving the AI ​​octopus "prosthetic" production line mentioned above into the body of the AI ​​octopus, and it doesn't even need to be "moved". It has very powerful coding capabilities. The only thing it lacks is the ability to execute code freely in a free environment. The previous GPT4o restricted the execution environment to a small closed space, and what Python-Use does is to provide AI octopus with a relatively free code execution capability to communicate with the surrounding digital world environment.

This is the story of the hero LGX saving the AI ​​octopus.

Comics generated by GPT-4o

Run away~ Let's continue~

So we naturally thought that the interpreter of the programming language became a very good combination point. The language interpreter naturally provides code execution capabilities and can be deployed in more scenarios, allowing the big model to better integrate the environment and open up the communication of various data. In addition, Python itself has a strong ecology and capabilities, and the big model itself is also very good at Python coding, so AIPython was born! We call it AiPy!

https://www.aipy.app/

The communication between Python interpreter and environment mainly depends on two things: API Calling + Packages Calling , which truly realizes the interconnection of all things + programmable everything! In the existing knowledge of big models, we basically have a lot of knowledge about API calls of general Internet services and the selection and calling knowledge of various commonly used Python packages. All the work can be completed by users entering the corresponding API key. Of course, for the latest API or private API, including the latest Python Packages, this requires the description of the API to allow the big model to complete the call! This replaces the previous Function Calling and Tools Calling. Because it is implemented by unified coding of big models, there is no need for the so-called standard protocol of MCP. Similarly, through the task understanding and planning capabilities of the big model itself, all tasks are fully executed through real-time coding, so Workflow is not needed, so those strange Agents are even less useful. If we have to talk about Agent, then Code becomes the only Agent, and the various big model clients used by users only need to install one AiPy.

So this is a brand new route: " No Agents, No MCP, NoWorkflow, No Clients... ", from this we believe that Python-Use is a new paradigm, which implements the new paradigm of " No Agents, Code is Agent. " Through Python use Data, Python use Browser, Python use Computer, Python use Anything...

Of course, it also has the ability to understand and plan tasks, code best practices, memory management and attention adjustment, and task reflection and feedback. We even believe that through the Python Use paradigm, various AI octopuses can work harmoniously for humans (similar to the "MOE" mode), and even allow AI octopuses to evolve themselves, not just hands and feet, but also promote the evolution of multimodal capabilities such as eyes, noses, and mouths! Ultimately achieving AGI is the real "intelligent body".

The Python-Use paradigm is a true AI Think Do, which means that AI can truly achieve "unity of knowledge and action". Through AI ThinkDo, we can make the big model understand tasks, plan tasks, execute tasks, and finally feedback the results in a more universal and unified way. Therefore, we believe that:

The real general AI Agent is NO Agents!

“The Model is the Product”

I guess many of you have read this article: "The Model is the Product" 

https://vintagedata.org/blog/posts/model-is-the-product

I have also introduced this in my previous article, and I agree with it very much. In this article, two modes are mentioned, which I prefer to call Agent 1.0 VS Agent 2.0:

The product of this Agent 1.0 era is a variety of "prosthetic" calling modes, which may be more intuitive when compared with the AiPy flowchart above.

This is Agent 2.0. In fact, when Computer Use came out, it belonged to this category. Of course, as mentioned above, the AI ​​octopus is driven by the visual mode. If the keyboard and mouse operations are simulated to complete the task, the current usability is still very low. The AiPython (Python-Use) we mentioned is written and executed through Python code to complete the Action and Feedback with the environment.

Then combining the above content, we have completed a perfect logical chain:

Model is product <--> Model is agent <--> No Agents, Code is Agent <--> Just Python-use <--> Freedom AI

So from now on: No Agents, Just Python-use!