I Turned My AI Coding Agents Into a Video Game

Author: Leon van Zyl · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: I Turned My AI Coding Agents Into a Video Game · Channel: Leon van Zyl · 16:55 · Watch on YouTube

Topics: AI agents, Task and workflow automation, Claude Code, Software engineering with AI

Summary

Leon van Zyl presents an open-source, 3D software factory where Claude Code, Codex, and Open Code agents work as a company, managed primarily through a CEO agent. The app connects projects to GitHub issues and pull requests, supports agent hiring and task assignment, includes QA review and optional automatic merging, and provides live project previews. He explains that he built the environment by iterating with Opus 5.5, then demonstrates installation and initial setup. The video ends with an invitation to try the project and join his Agentyc Labs community.

Key points

The 3D office represents a software factory populated by real coding-agent sessions.
Users can set a default coding harness, model, reasoning effort, and project folders or GitHub repositories.
A CEO agent coordinates projects, assigns tasks, manages workloads, and can hire additional agents.
The interface includes mobile access, a company overview, agent terminals, and a hiring waiting room.
Backlog items are GitHub issues assigned to agents according to their roles, with progress reflected in real time.
A QA agent reviews completed changes and can return tasks to developers; pull requests can be merged automatically after required checks pass.
Users can start a development server and preview the project, including its responsive mobile view, inside the app.
The creator says he built the project using Opus 5.5 and iterated on it through the CEO agent.
The project is open source and can be installed with one command; users can create a company and add existing or new projects.

Chapters

Start	Chapter	Summary
00:00	Introducing the 3D software factory	Leon introduces the 3D office and explains how it combines coding agents with a CEO-led company structure.
01:37	Touring the office and managing agents	He demonstrates settings, projects, the CEO agent, mobile messaging, and hiring developers.
04:52	Project floors, GitHub workflow, and live previews	The tour covers agent terminals, GitHub issues and pull requests, QA review, optional auto-merging, and in-app previews.
10:19	How the project was built	Leon describes using Opus 5.5 to create and iteratively improve the 3D software factory.
12:31	Installing and setting up the app	He shows the open-source installation process and how to configure a company, CEO, hiring preferences, and project.
14:11	Using the company and final invitation	Leon demonstrates the tutorial, hiring agents, and asking the CEO to create tasks, then invites viewers to try the project and join Agentyc Labs.

Quotes

What you’re seeing here is my software factory represented in 3D space. — 00:00

our primary contact is our CEO agent. — 01:05

we also get the report from that QA agent. — 08:10

this is completely open source. It’s free to use. — 10:19

You only have to run one single command. — 12:31

Concepts and entities

Claude Code, Codex, Paperclip, Opus 5.5, GitHub, Oracle, Oracle Agent Studio for Fusion Applications, Leon van Zyl, Agentyc Labs

Full transcript (original language)

00:00 So, who’s said that working with coding agents should be boring? What you’re seeing here is my software factory represented in 3D space. Each of these little workers sitting here are actual Claude Code or Codex agents. So, I’ve spent a lot of time building software factories, and I even have a dedicated video showing you how you can build your very own software factories. But, these basically come in two different flavors. Either you can have a software factory that hooks up directly to something like GitHub, and that usually means, you know, logging an issue, assigning it to a coding agent, and then one of the

00:32 agents in your coding swarm will implement the change and create a pull request. Or, of course, you can set up something like Paperclip, but this is yet another application or CLI tool that you have to self-host and manage. And I actually do have a dedicated video on Paperclip as well. But, this is just sort of a web interface or a terminal view where you define your company. You then set up your CEO agent, and you kind of manage everything through this web interface. It all works, but it’s not very sexy. So, instead, I decided to build something that’s less technical. You don’t have to manage GitHub, you

01:05 don’t have to worry about pull requests and logging issues yourself, but you still get full visibility of all of that if you want it. I also wanted to introduce sort of the org structure that you get from Paperclip, where we don’t have to interact with these individual agents ourselves. We can if we want to, but our primary contact is our CEO agent. The CEO agent can hire new agents. It can manage their workload and manage the agents across multiple projects. So, in this video, I am going to show you how I built all of this, and

01:37 you can actually use this application right now as well. I’ll show you how in this video. But first, let me give you a quick tour. On the first floor, we can access our very own office. From our terminal, we can access settings, and this is really where we can set our default coding harness. So, this does support Claude Code, Codex, and Open Code. We can select the model that we want to use, the default reasoning effort, and some other settings as well. From here, we can also set up new projects, hook up existing GitHub repos, or select an existing workspace or folder on our machines. At the moment,

02:11 I’ve got two projects set up, Glaze Club, which is sort of a donut shop, and Cube Farm, which is this very site project. Yes, I’ve been maintaining this project within its own 3D world. Then, if we head over to the CEO, this is our CEO. We can define the name and sort of the attributes and rules that this agent needs to follow, and on this board, we can see what this agent is currently busy with. We can also interact with these agents, so this will show us the actual terminal UI view, so the cloud code or Codex view for this specific

02:42 agent. We can also send messages to these agents from this little interface. And our CEO just responded as well. This CEO agent has access to all of our projects and it’s got a view of all of our different sub agents in the agent swarm. Now, this is similar to paperclip in the sense that we don’t have to interact with the individual agents ourselves. We can simply tell the CEO agent what we want to do, whether that’s building a new project or adding features or fixing bugs in existing projects, the CEO will create all the

03:13 issues and assign the tasks to these agents or potentially hire more agents if it needs to. Now, to interact with the CEO, we do not have to run to this guy directly. We can also pull up our mobile phone. From our mobile phone, we can interact with Morgan, which is our CEO. We can ask the CEO a ton of different things, like what is everyone working on, do we need anything new, plan new milestones, or we can just chat to it and ask it to hire more agents. In

03:44 fact, let’s do that now. Please hire more developers for our Cube Farm project. Let’s send this. and we can see Morgan is currently working. Also, if I just run to this board, where you can see that Morgan is indeed replying to me. So, we’ll just wait until Morgan’s hired more developers. From the mobile app, we can also view all of the hires, so we should see the new hires show up in a second. We also have our company view at a glance. Of course, agents can take a while to complete their work, so why not kill a little bit of time? Our

04:17 CEO already identified a few candidates, and these candidates are sitting in the waiting room. Now, what we can do is actually just walk up to them and click on any of these agents, and we can see exactly what their roles will be. So, we can even expand this job description and see exactly what this agent is responsible for. And if you’re happy, we can go ahead and click on hire, and that will move that agent to the actual project folder. Of course, we don’t have to interact with the agents directly. We can just go to our mobile phone, go to hires, and then hire the agent from here. Cool. Each of our projects are on

04:52 a separate floor within this 3D world. So, we can simply go to the elevator, click on this, and then let’s jump to one of these floors. This is the Glice Club floor. So, this is a 3D donut shop that we’re building in real time. Each of these agents are actual Cloud Code and CodeX instances working away. In fact, if we go to any of these agents, so let’s go to Ada, we can see in real time the actual terminal UI and agent browser. If we left click on this or press E, this will just sort of enlarge it so that we can see exactly what’s

05:25 happening in that terminal view. And if you do want to use this app like normal Cloud Code or CodeX, we can definitely go ahead and inject messages into the specific agent’s terminal. And on the right, we can see the actual website preview coming from the agent browser. This is all running in real time. But first, a quick word from our sponsor, Oracle. If you’ve ever built an AI agent that works great in a demo, you know what happens next. The moment a real business needs to use it, everything gets hard. Who can access what, who

05:58 approves changes, who keeps track of it all, and usually that’s where these projects get stuck. Oracle Agent Studio for Fusion Applications solves this in a smart way. Your agents run inside Oracle Fusion itself, so all the security, rules, and tracking that companies already trust, your agent gets that for free. And here’s the part I like, you build with the tools you already use every day. That could be VS Code, Cursor, the terminal, Git, and coding agents like Codex and Cloud Code. There’s also a public GitHub repo full

06:31 of templates, starter projects, and sample apps. So, you’re never starting from scratch. Simply go to github.com/oracle/fusionaistudio. If you’re already a Fusion customer, AI Agent Studio is included at no extra cost. And the GitHub repo is open to everyone. Go to oracle.com/applications/fusionai to check it out. The link is in the description. Thanks again to Oracle for sponsoring this video. Now, this agent I believe is using Cloud Code. Yes, it is.

07:05 And Grace is actually running Codex. So, if you open up its terminal window, we can see that this model or agent is using GPT-6 Astra AI on fast. Then also, we can run through the whiteboard. And from here, we can see all the items in the backlog. Terms, we can see who’s working on what. And how this works is when a new item is added to the backlog, these are actually GitHub issues. So, if you are familiar with using GitHub, this is what that looks like. So, we can see

07:35 for our Glaze Club project, we’ve got our GitHub issue. And this is currently being processed by our agent swarm. We can see it from this label over here. And these agents will add comments and update this issue as they’re working on it. These issues will automatically be assigned to the agents in the swarm based on their role. So, we can see that we have three agents working on this, and this is all updating in real time as you can see. And how this works is once an agent has actually completed its change, that change will be handed over to a QA agent that will test it and review the change. If the QA agent runs

08:10 into any issues, it will reassign this task back to the developer to fix any issues. So, what’s really cool is if you click on any of these PRs that have actually gone through review already, this will take us to GitHub as well to the actual pull request. From here, we can see exactly what was changed, but what’s really cool is we also get the report from that QA agent. So, we can see exactly what was reviewed, and this even includes test evidence like screenshots. So, this means that you can review the pull request and merge it yourself, but this app will also merge the pull requests automatically. So, you

08:44 don’t have to worry about that at all. You can disable that feature if you want, so there is this little auto merge from QA and check box. So, if you don’t want these to auto merge, you can simply uncheck this check box. So, how this app works is when a pull request is created, the QA agent will review it and then pass it or fail it, but auto merge will only happen once all the checks have passed. This includes your normal GitHub CI/CD checks, or if you’ve got Code Rabbit or Vercel checks as well, all of the checks need to pass before the PR

09:18 will be merged. In fact, we can see a card was moved to ready to merge. This means that this change actually passed QA testing, and in a second or two, PR 69 will automatically be merged with main. Now, this is all cool. We can see the agents work away, and we can see what they’re all busy with. But, how do we preview the actual app? Well, what’s really cool is we don’t have to leave this app to see our actual project in action. We can go to this monitor over here. We can start the dev server really easily. So, if I clicking on this

09:48 button, this just started the dev server for this website, and now we’ve got a live preview of the actual website. So, as these agents are making changes to the website, we’ll see the changes reflect immediately. We can also click on the phone view to view the responsive design of this website as well. So, everything is running and being changed in real time by this agent swarm. And like I mentioned, there are so many ways to kill time while we wait for the agents to do the changes. You know, we’ve got 3D objects that we can

10:19 interact with. We can try to shoot this ball through the hoop. Everything just works. And of course, we’ve got this little Nerf gun as well, so we can go ahead and maybe shoot at some of our agents just to see their reactions. It would have been cool if some of these agents would actually shoot back at you. Also, this is completely open source. It’s free to use. So, let me show you how I built this and how you can start using it today. So, first things first, I’m sure a lot of you are super curious as to how I actually built this. Now,

10:50 firstly, let’s talk about the model that I used. Opus 5.5 is stupidly good at doing 3D work, especially in building game environments. So, for this entire project, I only used Opus 5.5 and I mostly ran the model on extra high. And this really is the prompt that I sent to Opus. So, I just kind of told it to build a browser-based 3D environment. The graphics should sort of be cartoonish and it’s from a first-person view. And I just initially told it to use Cloth Cat and the Cloth Agent SDK and use my subscription to build this

11:25 agent swarm. I then just said that I do want this to act like a software factory where it’s creating GitHub issues and pull requests. And you know, each agent’s got its very own code code session effectively. So, I’m not going to read through all of this. You’re more than welcome to read it yourself. But, the main elements like the Kanban board was already mentioned, and that was really it. So, that prompt gave me about 50% of what you’re seeing here. The agents could sit behind these desks, and obviously we didn’t have things like the toys and the balls. The whiteboard was there. It was missing the monitor, so we

11:59 couldn’t actually preview the app. But, you know, a lot of the items were there already. Then, what I did was actually open the project, so this very same project in this 3D world. And at that point, I just kind of started asking our CEO agent to add more things like this little monitor that we can use to preview the app in real time, and you know, just support for things like Codex and open code. So, you know, I just kind of played around in this environment and asked the agent to keep iterating on it. So, this is showing you sort of the cost

12:31 and the time it took for this initial session. I’ll just open up the stats. And here you can see the entire bolt used Opus 5.5. It ran for about 1 hour and 16 minutes, and only cost like $35. If you would like to try this yourself, or even fork this project and make it your own, this is completely open source. I’ll link you to the GitHub repo in the description of this video, and this is really simple to set up as well. You only have to run one single command. So, I can just copy this command. And by the way, if you do want to support this

13:04 project, then consider giving it a little star on GitHub. And so, all you have to do is run that command in your terminal or command prompt or PowerShell, whatever you want to use. And that should automatically install all the dependencies and open the browser for you. So, all we have to do then is give it our name and our company name. We can go to next. Then, we can set up our CEO. I’ll just call mine Luna. I’ll set it to a she and maybe change the color. Then we can decide on how hiring should work. Our CEO will

13:38 automatically try to hire sub agents for our swarm. Now you basically have two options. We can decide to approve the hires manually or just let Luna hire new agents automatically. I’ll just review the applicants myself. And now we can set up our very first project. So you do have a couple of options here. We can build a new project from scratch or we can select an existing project from our PC or we can hook up a GitHub repository. I’ll just select the folder on my PC. So from my projects folder, I’ll select Glice Club and let’s add

14:11 this as a new floor. We can add as many projects as we want. Now our company is open for business. So we can enter the office. This will also give you this quick tutorial. So we can press P to open up our phone. And this is where we can interact with Luna, our CEO. We can view new hires. We can see our entire company at a glance. And of course we can play games like the snake clone, Tetris. We even have this little desk pet game and you kind of have to feed it

14:42 and play with it and keep its mood up. If you ever want to add more projects, really all you have to do is go to your machine and [clears throat] then just add a new project from your machine, a GitHub repo or build something new. Let’s visit our CEO. And you can see what our CEO is doing by pressing E or the left mouse button. So at the moment our CEO is trying to hire new candidates. So we’ll just wait for that step to complete. I’m just going to skip the tour because I want you to experience this yourself. So what I would recommend is just wait for your CEO agent to hire new employees because

15:17 at the moment we don’t have any agents in the swarm. Once we have our new hires, what you can do is ask your agent or your CEO to create new tasks. And it’s really as easy as saying, “Hey, for the Glice Cloud project, this can you add more tasks? So, improve the visual design, maybe the variety of donuts.” Cool, I’m going to send this to our CEO. And by the way, I can see the new hires are ready. We should see them and in fact we do. We can see them sitting in the lobby. So, I’m just going to hire Ada, Linus, Grace, and Alan. Now, our

15:52 project shouldn’t look that empty. And of course, we’re waiting for the CEO agent just to add more tasks. Let’s go down to our actual project folder then. And we can see all of our coding agents sitting over here. And because this whiteboard is actually hooked up to a real GitHub repo, I can still see all of the issues and the PRs from my previous session. So, let me know down in the comments what you think about this project and please try it yourself. I would love to get your feedback. If you would like to learn how to use coding

16:22 agents to build cool projects like this, then consider joining my school community Agentyc Labs. I’ve got several classrooms and hours worth of videos showing you how to use coding agents effectively. From Agentyc coding fundamentals and setting up your development environments correctly to creating design systems, animations, and we’ve got projects that you can follow along with. We also have live weekly sessions. So, every Tuesday we’ve got a very casual coffee session where we can chat about anything. It can be career advice, issues, new model releases, whatever. And every Thursday we’ve got

16:55 deep dive sessions. Like tomorrow we’re actually discussing Jeff from TypeSafe AI. I’ll leave a link to Agentyc Labs in the description of this video.