Transcription
Hello everyone. You are on the AI Dialogs channel. And a few days ago, OpenAI released Codex App, an application for working with code agents. This is not a new application, but rather a rethinking of how to work with code agents. And most importantly, according to Sam Altman, for the next 2 months, this application will be absolutely free to use, even on Free and GO tariffs, as well as use increased limits on all paid tariff plans. And what is important, under the hood is the powerful GPT52 Codex model, and it is definitely worth trying. OpenAI calls the interface reimagining a command center for working with code agents. But how convenient is it, or maybe it's a beautiful shell with a lot of real problems? Today we will figure it out, go through all the main functions, and, of course, see how to work with this application in practice. And you will be able to decide for yourself whether it is worth it or not. Let's start with the press release. A powerful interface designed specifically to work with multiple agents simultaneously, run tasks in parallel, and work with agents that can perform their tasks for a sufficiently long time. Well, they say that Codex is, of course, not a new tool. It's almost a year old, changing, firstly, how software is created and who can create it. And they say that the creation process is already shifting from pair programming with an agent to, in fact, coordination, management, and delegation throughout the entire lifecycle. And they claim that Codex is like a real command center for agents. Like the reasons why they decided to rethink. In principle, since Codex appeared in April 2025 as a tool for cloud-based execution of various tasks, how agents work in general, what tasks they can solve, how to interact with them, has significantly changed. In their opinion, they believe that this process has actually shifted from direct interaction to asynchronous delegation and management on a large scale. And they say that existing integrated development environments, for example, Cursor, or terminal agents. Hello, Code. They are already not very adapted to such an interaction model. And so, they are creating an interface that is more suitable for this. Command center, all that. Well, we will look at it. It flashes on the screens here. Let's go through the key features. Well, firstly, working with agents in parallel. They provide the ability to launch agents, well, they actually call these chats with agents "threads," which are organized by projects, with which you can easily work in parallel without losing context. Well, and plus, they support, of course, working with GTWes, which allows agents to work in parallel without interfering with each other or conflicting in terms of source code. And in general, it's probably worth noting that parallel work with agents is not a new feature at all, and it is present in most tools in one way or another. And we will probably look at that too. They say that their tool uses and supports skills, which is the ability to extend agent behavior using documentation, resources, and even source code. This should be noted as an absolutely powerful and promising capability, but it is not an OpenAI invention. This is already a standard developed by Anthropic, like the previous popular MCP, and which, in principle, all manufacturers have supported quite quickly this time and have implemented support in their tools. That is, this is a very useful technology. It can bring the use of such tools to a completely new level. Well, and they give an example that they supply their product with a number of ready-made skills, including a skill for creating their own skills. And this, in general, will be a convenient and in-demand story. Well, and they give an example here that, say, from one prompt, we used a skill designed for creating games. And we made several game variants with different token consumption, from 7 million to, more precisely, from 60,000 to 7 million tokens. You can do many different things with skills, work with designs, manage projects, deploy, for example, generate images. They present their skill for image generation using ChatGPT. And much more. Separately, it is worth noting the cool, absolutely in-demand, useful skills for generating various templates from, for example, text documents, in formats like Markdown, and then creating office spreadsheets or docx documents. The last killer feature is so-called automation. And they say that you can delegate repetitive work in this way. Create automation, it will run on a schedule and can use all the functions of the code agent you are working with. That is, it can use skills, etc. And as an example, they say that in this way, in principle, we check on a daily basis whether our builds have broken, look for bugs, generate release notes, and so on. But in the context of local, yes, this agent is installed on my computer, and the automation function seems a bit far-fetched, because, well, it's hard for me to imagine a task that, for example, an agent on my computer would start doing on Monday at 9 AM on a schedule. That is, rather, these things are more like something that should be done somewhere in the cloud, and not depend on being run locally. And secondly, these are still some truly corporate, possibly, services, CI/CD can run tasks on a schedule. For local use, you still need to find a task. Rather, it's an execution of actions on demand. That is, I want, for example, to send my changes to the repository. Well, why not do a code review locally before that, look for bugs, analyze the code using some hooks. And that's all, in general. It's better to see once than to hear seven times, as they say. Let's see how it looks in practice. I remind you that, unfortunately, for now, it is only available for macOS users. Before us is the interface. On the left side, we have, in fact, the declared two killer features of automation that can be created. And here is a set of ready-made automations. Well, it's still beta, you can create your own automation, name, prompt, on which project to do it, actually, what schedule, regularly or at some interval. In general, an interesting function, but probably you need to think of something else to use it for. I already mentioned skills, the ability that definitely needs to be applied in your projects. And probably, in addition to publicly available skills, you need to create and package your own workflows into such extension packages. This is an endless area for increasing the efficiency of your activities. That is, all your workflows can certainly be made repeatable and reliable. Dialogues with agents to solve a particular task. Threads are organized by projects, yes, you can either create a project or add an existing one. Let's add my existing project and try to do something with it. I've already experimented a bit here, but that's not a problem. We select the project. Here it is selected. And, by the way, among the useful functions, they have also implemented the not-so-new voice input feature. We will definitely test it now. What can be done here? Of course, you can add images or files to the context. You can use the context of the integrated development environment. Without it, obviously, it won't work, but it's not here. You can use planning or execution mode. You can choose models. In this case, one model with different levels of reasoning. You can also use various commands available by slash. Ready code review, using MCP, enabling plan mode, or, for example, creating skills or installing existing skills. Regarding further work with code. There is an option to simply use a local project. If we are going to run agents in parallel, then it is advisable to use the Work 3 approach, so that agents do not interfere with each other. And we can open and edit the results or individual files in one of the installed on the workstation. I have many different ones installed here, you can choose any. Well, by default, of course, Cursor. Obviously, it's clear that you can't do without an IDE. Well, and here you can also see the menu item for working, in fact, with the repository, sending changes to a remote or local repository, creating a pull request. And there is also a panel where you can see the changes that have been made, and there is actually an option to call the console. Basically, it's quite minimalistic, but let's try to do something. So we have our landing page. It's a regular web application, with several pages. And we will try to improve one of the pages. Specifically, on the main page, I want a new section to appear. Let's try this. On the main page, add a new section, please. This section will be for media materials with links to our articles in open sources. And add two elements there, two articles from Habr. The links to the articles themselves, which will be displayed there, should be configured. So, well, it transcribed it, respectively, we didn't provide anything to its context. Well, it started looking at the project structure itself, found the necessary files, saw what layouts and pages there are. Well, and now, obviously, it will find a similar template and add a new section. While we wait, obviously, to check, we will want to run our application. And here, in general, is probably the easiest standard way. You can, in principle, ask in the chat to do everything we need. This is the standard path. There is also a button provided with which you can run the project. Well, like this. The first elements, ripped from the IDE. Well, we know that, for example, our project starts with the command Make dev. And so, well, we don't want to run anything yet, yes, let's stop. But what does it do? It simply runs this command in the terminal. In principle, we could have done it ourselves. Let's see what it did here. It added config. If we want to see it directly in the chat, we can click this button. We can also want to see it in an external editor. We can click on this file. It will open, yes, with the result, in this case, Cursor did it. Everything is done. Let's see how it looks. In general, a button "review Changes" appeared for a fraction of a second, but now it's gone. Yes. The only thing is that we can cancel here, but we have a right panel that is designed to track these changes. In fact, and they show in their materials that you can leave some comments here, based on which the agent will then work. And we can carefully look at everything it has done. In principle, this mode can be called the review mode, and here is the final part, that we can commit all these changes, and then send them to the repository. But let's check first, for this we will run it and see what happens. The terminal starts. It should be noted that there is only one terminal here, which, in my personal opinion, is a big drawback. Yes, I cannot do anything in parallel here. Well, let's see what it did anyway. So, here is our landing page. We scroll down. And it added the media section, publications, articles from open sources, and added, respectively, two blocks of links to some articles from the config, but I didn't give them, obviously, so it made some arbitrary ones. Further, respectively, we looked at these changes. Okay, we checked everything there, and we can probably commit all these changes or commit and push immediately, or we can commit and create a pull request. That's basically it. That is, well, it's clear that we can open this project in, and that's where all the possibilities end. That is, yes, indeed, we can send tasks to agents in such a dialog mode, they will work in parallel here. But, there's nothing more. And I must say, one of the main drawbacks, at least for now, is that I cannot actually add any context here. That is, the project tree is hidden from me. And we generally consider it good practice. In the industry, the approach of going from some documentation, from a specification, Spec-Driven Development approach, is also considered good practice. And it becomes more difficult to do it here. This interface is quite interesting, but I must say that it is not new. For example, I want to show how the interface for working with agents looks, for example, Cursor, that is, in an integrated environment with agent support. Well, in general, there are parallels. What's interesting is that to run and check, we need to open a browser. And if we want to do something with the console, then, in general, only one console is available to us for now. So, what do I want to say in conclusion? A command center for working with agents sounds very cool and promising. Yes. OpenAI hid, moved away all the complexities inherent in IDEs. But in reality, it seems to me personally that they hid not the complexities, but they hid and removed part of the functionality, without which, in fact, working with any project becomes quite inconvenient. But this tool is definitely worth attention, to try to use it, especially considering that it can be done absolutely for free. And from the perspective of competitors, opinions are divided in the industry, in the community of developers and those who use code agents. Half said it's a breakthrough, a step forward, a very powerful shift. The other half said that OpenAI is catching up with Anthropic, Codex is trying to catch up with Codelabs, and that Codelabs' interface is generally more powerful, and workflows in it are more advanced. And here I want to raise an important issue. Code agent manufacturers compete, new products or powerful updates to existing ones appear every week, every month, and tools change, but the problems that specialists face when working with code agents do not. Many things that we observe regularly. The code generated by agents is difficult to maintain. Documentation is not synchronized. Tasks left by agents are poorly decomposed. It is impossible to maintain projects generated using code agents. Any attempt to make changes after some time simply fails. That is why we created the Smart Coding program. The next session of this program starts next week, February 13th. And in this program, we teach not how to work with a specific tool, but how to build a system that will allow you to get reliable, high-quality results using any code agent. And among other things, in addition to methodology and system, you will be able to create a whole series of projects yourself, based on large language models. It is a good start for a wide audience to start creating projects independently. Go to the website, check out the program, sign up. I also remind you that we have a Telegram channel where we publish more useful materials. That's all from me. Thank you all. Goodbye everyone.