https://pronetic.geeknetic.es/Guia/3387/Probamos-OK-Computer-la-IA-agentica-de-Kimi-AI-Merece-realmente-la-pena.html

OK Computer is the agent of Kimi AI

Although there is no standardized definition to describe what agentic AI is, almost all experts agree on one thing. An artificial intelligence agent must be able to execute tasks on its own, without human intervention. Kimi AI is a Chinese platform that has a function called OK Computer. Is he an agent? Is it a vibe coding tool? Well, a little of everything.

With OK Computer mode you will be able to create applications, websites, presentations and many other digital resources. This is a fairly solid proposal, as you can see in the examples that I leave below. The thing is that she is still making mistakes and is quite far from perfection.

So this is not a guide where I teach you how to create a perfect business application in just 30 seconds. Rather, Kimi’s OK Computer mode (like any other AI service, really) is a great tool for coding faster, improving workflows in your company, or generating content in less time. Don’t expect a revolution, but expect substantial time savings when launching projects or learning to program.

By the way, to activate OK Computer mode, you just have to enter kimi.comlog in with your account or create a new one and click on the button that you will see just below the text field. Once you do, you can start experimenting with Kimi’s agentic AI.

Example 1: create a product listing by consulting Amazon

After activating the mode OK ComputerThe first thing I tried is this:

I need to make changes to my setup. So, go to Amazon and look for the best wired mechanical keyboards. Keep the three that have the best reviews and save them for me in a list. Then search again to find three good mini PCs, mainly with AMD processor. Finally, find three standing desks with at least 160 cm width.

As you see, there are at least four tasks here. Kimi has quickly identified them:

Find the keyboards.
Search for mini PCs.
Find the desktops.
Create a web page with the list of products.

While it is true that I have not asked for the latter, it may be useful. What has been the result? With a single interaction, I have obtained this:

Although the product list is cute, the truth is that it does not include any link to Amazon. In any case, the selection of products is real and can be very useful when starting to research new products for a setup.

Example 2: Develop a Trello-like Kanban board

For this second example I have used the following prompt:

You are an expert web developer, specialized in creating high-fidelity, intuitive user interfaces, and your task is to develop a task management website that rigorously implements the Kanban methodology, faithfully replicating the Trello interaction experience. The site should include a customizable dashboard background, a clear navigation header, and horizontally arranged lists, card lists, and subtask behaviors (Drag & Drop) for moving and reordering. Each list and card must support online editing on any card, with a quality editor that allows you to modify the title, description, add color labels and manage checklists. Prioritize a responsive, clean and minimalist design, with slightly rounded corners on key elements. Provides the core HTML structure code and basic CSS/layout for the horizontal layout and card styling, and includes a brief description of the JavaScript interactions needed to achieve the Trello experience.

As you can see, the order is quite extensive and includes a lot of details. After a few minutes, Kimi has completed the task. In this case, what you have done is package the necessary files for the application to work directly in the browser.

Just one detail in this step: I was not able to download all the elements at once, since the ZIP it provided came empty. I had to download the files one by one to be able to run them on my PC. With a single prompt I have obtained this application:

Again, it’s a pretty striking result, but with some problems:

The changes are not saved, but are lost when the page is refreshed.
It is not possible to sort the columns.
After adding an item, no more cards can be moved.
It would be interesting if there was a context menu with actions for the cards.

In light of these limitations, I have asked OK Computer to resolve these issues. And the truth is that he has done it. The second version of the Kanban board already had a context menu, column reordering and a database in markdown format.

But all that glitters is not gold. While Kimi has eliminated and improved the functionalities from a first version, he has added new errors. Without a doubt, this can become a real hell for someone who wants to launch a serious project.

There is still a way to go in agentic AI

Write a prompt extensive with a lot of tasks for the AI and having it execute them sounds good. It looks like this will be the next big revolution, even if 2025 ultimately isn’t its year. The thing is that reality is much more stubborn. Language models have serious limitations when it comes to completing complex tasks, much less when they are accumulated in the same prompt.

The same can be said of the tools of vibe coding. If you have no idea about code, services like Kimi may help you do something small, a necessary utility for your daily life. But the moment you want to scale the project, The limitations of AI will quickly appear. And if you also lack the knowledge to fix the model code, say goodbye to continuing.

Therefore, all this does not mean that authentic modes like OK Computer are useless, but rather that they have important limitations. For example, they can help you with specific tasks that do not require scalability:

Create a tool to make specific calculations for your work.
Develop an article search engine on a specific website.
Add functions to existing code of a website.
Prepare a presentation from data.

Right now, I think we have to wait to see if the models continue to improve over time. AI agents They have to mature to be taken into consideration. They need to be more precise, better understand what is expected of them and reduce their randomness as much as possible. Even when an agent works well for a specific task, it is susceptible to breaking if the model is updated to a new version. Who wants this in a business environment?