I'm the director; the agents are the crew

Hey there! In the welcome post I said I’d write more about the tools I’m using to build this thing, so here we go.

First, just to put it out there, this all still feels like magic. I am a huge nerd but I am not a real engineer, and suddenly I can build serious, deep, and useful sites, apps, tools, artifacts, you name it, and all about 50 times as fast as an ancient human from 2024 could. How? Robots.

(Note: I am being silly calling AI agents “robots”; I get that robots are a different thing in the common nomenclature, but I like calling them robots so that’s what I’m gonna do. Hopefully the context is obvious enough that no one gets confused.)

So, as I’m sure you’ve heard, these robots are getting REALLY good: top software engineers went from “I write code” to “I only review code” to “I don’t look at code anymore” in about a quarter, which is mindblowing. And [insert Terminator/SKYNET reference here] now the robots can create other robots to do work for them. And last week my favorite robot (Claude) got a huge upgrade and can dyanmically spin up dozens of specialized robots for various purposes in order to build a lot of things really quickly (or to research, investigate, analyze, write, use tools, connect to other robots, etc). It’s called “ultracode” and it is no joke; Opus 4.8 in ultracode mode is by far the most impressive thing I’ve ever seen from an AI lab.

So that’s my primary tool. I dabble in other tools, and the Codex desktop app is, in my opinion on this particular date, the best of the agentic IDE’s yet. And GPT 5.5 Extra High is very impressive and capable. But I keep coming back to Claude Code in the terminal because I find the interface more delightful and I preferred its harness even before ultracode. Now in this moment, with ultracode, I’m not aware of anything that can come anywhere close to what I can do with Claude. To be fair, lots of brilliant people there have put out open source harnesses that are super sophisticated and maybe better than what I’ve got; I mostly haven’t used them though, because I find my flow with Claude to be so effective I’m not compelled to seek out outside harnesses).

What’s that flow? Generally speaking: Establish intent (let it know what I’m working on or trying to achieve), then ask it to scaffold deep research (it customizes this every time and always comes up with areas to reearch that I didn’t think of), then present it’s findings with decisions needed and recommendations, then iterate on a design, then build an implementation plan, then iterate on build > validate > reject/update/continue. It’s not rocket science and like I said other people have smarter approaches, but this works for me. Even for some little personal app that I’ll take from first idea to done in half an hour, I run through these steps. I have found that the mostimportant out of all of them is the research; that way the robots have that corpus as a foundation instead of whatever generic, high-level knowledge they have from their training data.

And here’s the fun thing about Daedalus that arose organically: Taking a step back, a key tenet of my entire journey for the last year has been to, whenever a task arises, DON’T DO IT; figure out how to get a robot to do it. This applies to essentially everything, but is also of course necessary if I don’t know how to do the thing. So, I want to make a game, but I don’t really know how, so ask a robot how. The robot tells me I need a harness for that. Ok, how do I make a harness for that? Turns out to build an entire game would take a hell of a harness, and as I dove into learning about that, a light went off: If I can figure out how to get robots to build a game, then couldn’t other people use those robots too?

So here we are in a bit of a meta proving ground: I am trying to use a crew of robots to build a crew of robots that can build a game. In other words, I’m designing and building a studio team for me to build Daedalus, which is a studio team for me and you to build games. Of course that’s a lot harder than just designing a harness that could build the game I want to build; getting a swarm of robots to build a colony sim is orders of magnitude simpler than getting a swarm of robots to build a platform a human user could use to build a colony sim or an action RPG or a platformer or an arcade or a 4X or whatever else the human can think of. But if I can do it, I’ll be the happiest man alive because 1) I will be able to use it to build things that entertain people, and 2) I will be able to help more members of the community to bring the “make a living from making games” dream to reality.

While these new tools are magical, this work still isn’t easy, which is why nothing like it exists yet. The agents are astonishingly capable and fast, but they’ll also happily sprint in the wrong direction if you let them. The hard part, the part that’s still entirely human (for now), is knowing what you actually want, catching the robots when they drift, and having the judgment and opinionated taste to say “I hear you but that’s not the direction I want to go; lets do…” or “wait, why did you do that? The goal is to x, so we shouldn’t ever do y - save that to your memory so it doesn’t happen again” . So a lot of my days are spent writing specs, watching the robots thoughts to make sure it doesn’t do anything crazy, and reviewing, reviewing, reviewing. The leverage provided by these things is incredible, but they still require a Director in order to make something good. And I love being the Director.

With Daedalus, we will ALL get to be the Director of a commercial-grade game studio in our pocket (or Mac mini or whatever).

I am trying to follow the industry and as far as I can tell there is no single platform that can take you from idea to checks from Steam, without writing any code, drawing any sprites, or recording any music. From what I’ve seen there’s nothing even close. And honestly there’s a good chance I simply won’t be able to do this. But I’m gonna pour my heart into it until it someone else does it well (and maybe still keep pouring my heart into it once they do).

So that’s the meta-story of this whole devlog: a guy who can’t really code, directing a studio of agents to build a tool that lets people who can’t really code direct a studio of agents. It’s turtles all the way down, and I find that genuinely delightful.

More soon, including, before too long, the thing actually doing something. I can’t wait to show you.

Cheers!