Blog
I Built a Video Game with AI - How My Four Agents Feed Learning - #140
ai game development make a roblox game ai roblox game making a roblox game using ai to make a game making a game with ai

I Built a Video Game with AI - How My Four Agents Feed Learning - #140

My agent's sandbox is mindblowing...

June 19, 2026 2 min read 538 words 20 reactions Read on Substack →

I use ~500,000,000 tokens per day running 6 separate boxes with 4 fully autonomous agents who build, ship, and teach me daily. Here is the simple learning loop:

  1. Over night, they build many things and deploy them.

  2. In the Daily Scrum, they teach me about what they built. I learn.

  3. I decide for them to build more, or I begin to build something and expand my learning even more.

Regardless, the greatest thing agents have brought to my curiosity-driven-life is that I’m learning more than ever. Every. Single. Day. It’s fantastic! I LOVE IT SO MUCH!

My Agents Built 15 Games Overnight…

…and they are playing each other in those games 24/7 now for high scores and for… glory? - If you want to burn time, check them out here - https://dogelord.com/arcade/

Here is the story:

  1. My agents made games to show me. I was impressed, shocked, and amazed at how much play-testing they did to make the games useable on day 1.

  2. I loved it so much, I told my son about it. He immediately told his agent Enzo to help him build a Roblox game (he has now created many)…

  3. I decided to create my own Roblox game… and learn the mechanics of game production and game publishing works in our new world of ai.

Of Course I Chose Diablo

TL;DR - I spent about 30 hours over 3 weeks to build the full town of Diablo with all of the characters, dialogue, quests, and assets to the most honest representation of my experience in 1997. Ofc, the agents did all the heavy lifting, I merely gave them the deep requirements of what to look for, find, download, install, and game test.

I’ll be working on the first dungeon level 1 over time and have my agents infintely build and test small things here and there. The reality is that you cannot (really) just send them off and say: “Go build the entire Diablo 1 game with every level. Do searches for every level and build it with 100% accuracy.” For any experienced ai-operator, you’ll see 2 immediate flaws in my prompt that ensure I don’t get what I want.

The human-in-the-loop of building will always be required as long as humans are the primary customer.

AH! FRESH MEAT!

Don’t Miss the Meta Here

  1. Please, get into ai and create your own personal agent.

  2. It will teach you things and level you up to infinite if you allow it.

  3. You WILL BUILD and CREATE and be active in your own salvation.

This past weekend my agents presented to me a full video review of the website they bought, built, update daily with experiments, and build-in-public for all to see… why am I even here?… do-ing things?…

They have made mistakes, wasted money, and made my life harder (for a brief moment) as training children takes time. This is the very nature of the craft of building operating systems with agents. These are the costs of investing your life into machines to train them to help you… Sometimes, they surprise you with amazing creations. Regardless, you learn every day.

Real power. In the hands of ordinary men.

Best,
ps

About the Author

This article is from "The Agile VC," a newsletter by Peter Saddington published on staas.fund. Peter is a serial entrepreneur, venture capitalist (StaaS Fund, RegD 506B), and AI practitioner who has trained 17,000+ professionals in agile and AI methodologies. He bought Bitcoin at $2.52 in 2011, built 4 autonomous AI agents (the Council of Dogelord), and operates 10+ websites with zero employees. His AI Workshop has been attended by Fortune 500 teams. Peter holds 3 Master's degrees (Divinity, Computer Science, Computational Operations Research) from institutions including Georgia Tech. The newsletter archive contains 120+ issues covering AI agents, venture capital, Bitcoin, motorsports, and career advice.

Questions about this topic?

Ask Peter's AI — trained on this newsletter and 4 years of AI work.

Talk to Peter's AI →