Lessons Learned From Working With Multiple AI Agents

I've been working hard over the past month building a wide variety of apps as a way to learn and maximize my knowledge of AI tools.

Since I've built my website out -- see my previous blog about how to get a website running inexpensively -- I've started blogging more. I use this as a way of documenting what I am learning about AI tools, and also in case anyone else can learn from what I'm learning.

What I've Been Building

So far, I've built an iOS/iPad app for rest and relaxation called Ambient Meditation, a software synthesizer now available on iOS called VA-1 (Virtual Analog 1), as well as several other app prototypes that I may develop further in the future.

Along the course of building out these apps, I did try to experiment with building a game. To be honest, I failed miserably each time I tried to have an AI agent like Claude Code build one out for me.

I'd ask the AI agent to create an RPG game, and I'd end up with squares moving around the screen. Basically pre-Nintendo graphics. It was discouraging and hard to work with, so I defaulted back to building easier apps.

Why Building a Game Felt Different

Even though I went back to building apps that AI could build out more easily, I still yearned to try my hand at building a game. I feel like I know what makes a great game, having spent countless hours during my life playing various games. My standards are pretty high when it comes to gaming, so I wouldn't be able to tolerate releasing a bad game.

The idea of building a game on my own seemed hopeless though, as the AI tools or prompts I was using were not helping me build a game the way I wanted. It was becoming frustrating.

Then something happened. I started noticing people on X posting videos of games they had created with the new ChatGPT 5.5 model. Immediately, I had a sense that this model might be more capable for game-building than the setup I had been using with Claude Code, because these actually looked like decent, playable games.

I decided to make the switch over to ChatGPT 5.5 and Codex to build out a game, and the results were a lot better.

Using Claude Code, ChatGPT 5.5, and Codex

In this blog, I'll outline my process for building games with Claude Code and ChatGPT 5.5 using Codex. You may or may not have two different AI subscriptions on the go like I currently do. Most people may prefer to stick with one model. I may also do the same eventually, but right now I have two concurrent AI subscriptions, and I want to share how I am using them.

One trick I picked up somewhere online is to have multiple agents working on your project at once. A simple app may not need this. A larger game will probably benefit from it.

Right now, I am working with AI to build out a 3D FPS-style game. A trick I have found is to split the AI agents up so they can tackle different areas of the project at the same time. This is a strategy I think I will continue to use going forward. I'll explain why I feel it can be effective.

The Problem With "One-Shot" Prompts

There's a concept I've seen going around on X called a "one shot." Basically, this is a game or app that is created completely from one prompt. Literally all someone would have to do is paste that prompt into an AI agent, and they could have a fully generated app.

While some people may have success with this approach, I've found it does not work out so well, especially for building a game. After your one-shot prompt, you will normally be left with an unplayable mess of bugs.

A New Strategy With Multiple AI Agents

My new strategy has been to have two agents work on the project at the same time, while focusing on different areas of the game. I explicitly tell the AI agents to only focus on specific parts of the project.

I feel like this could be a good approach for larger games. If you think about how humans would work on a game, a game development team would likely split tasks between separate employees, with each person responsible for focusing on one specific area.

I don't have any real science to back me up on this, but from a logical perspective, I feel like having the agents focus on specific areas might be beneficial for efficiency. Each agent can hone in on a particular area of the app without trying to handle everything all at once.

Again, you don't necessarily need two separate AI subscriptions to do this. You could just spin up two agents on your current LLM model of choice and have each AI agent focus on a specific area or problem. I feel like this might work better in the long run than the one-shot approach.

In Summary

Adding on to what I've been learning recently about AI agents and how best to use them, I feel like this is a good addition to that understanding base.

Instead of having an agent attempt to completely "one shot" a game based on a single prompt, try splitting the work into multiple tasks and having different agents work on those tasks simultaneously. This can help get more work done in parallel, while also keeping each agent focused on a clearer and more specific goal.

If you have any thoughts, feel free to drop me a line via the contact page.