The 4-Step Agentic Engineering Workflow I Use to Ship Real Apps | Ras Mic

Michael Shimeles (0:00)

Software Factory, as it's explained and broadly known, I'm just not a fan of, right? Because one bad, wrong assumption in the beginning just destroys everything. I'm not at a point where I trust the agent. Most models, your Fables, even Astra, usually the first PR they make is a four out of five. This is essentially the crux of the workflow. Isolate, new work tree, write code to a specific standard, prove that a bug exists, even though this is supposed to be Software Factory, it's very much human in the loop. I am testing my features.

Peter Yang (0:33)

Hey everyone, today I'm really excited to welcome Michael, a full stack engineer, who also has a popular YouTube channel where he teaches 100,000 people how to code with AI. And today Michael is going to show us how he set up his Software Factory and the agentic engineering system. Personally, I'm a little bit skeptical of Software Factories, because when I try to do stuff like that, it tends to stopify everything. But I think Michael has found the way. So really excited. So welcome Michael.

Michael Shimeles (0:55)

Yeah, thank you for having me, Peter. I'm really excited. And I agree with you too. I think the general term and what you hear of a Software Factory, it usually is a slob cannon.

I have a different version, a different, sorry, definition of what a Software Factory is. And I'm happy to share that. To me, a Software Factory is more about your workflow, skills and domain knowledge, and sort of packaging that in either an agents.md cloud.md file and a bunch of skills, right? You know, going back to core engineering principles, like many, I've worked with many engineers. We all kind of work differently. And to me, a Software Factory is being able to take that domain knowledge and sort of have a system where you can easily go through the flow. So a Software Factory, in my humble, respectful opinion, I know there's a lot of startups that are doing this, is model and harness agnostic, right? So it doesn't matter what harness you use or what model you use. Obviously, you want to use the best model. That being Astra right now, in my humble opinion, all the Fable 5.1 is great at writing code. That's what a Software Factory is. And I think I'll probably give you like a brief breakdown of how mine works. It's just exactly how I would write code normally. I have, I'll just delete some of these examples so we can walk through them together. I have basically one agents.mb file and a couple skills. And I think the best way to explain this is sort of show my agents.mb file.

And my agents.mb file, the purpose of it is it just explains my workflow. It doesn't really talk about how the code base is structured and stuff like that, because the models are so good now, they can use their read tools and figure out how things are structured. But what my agents.mb file sort of explains how I want my agent to work, right? And there's essentially four steps.

The first step is to isolate, right? So whenever I go and I prompt for a new feature, or I'm in like discovery mode, or I'm telling the agent, okay, I want you to build, you know, analytics, or I want you to improve email tracking, or I want you to, you know, I want us to have, an iMessage service run on my MacBook that's on my network, right? So whenever we're working on a new feature, or we're fixing a bug that exists, just a fresh, fresh new work, what is going to happen is this skill called new feature is going to be called. And basically what new feature does is it branches off a Git work tree off of the main branch, right? So if I were to think of this is my project, right? This is like the start and this is the finish and this is my main branch. Every time I'm working on a new feature, what essentially happens is a new Git work tree is created branching off me. And what this allowed me to do is I'm usually working on multiple features at a time. Agents have allowed for me to explore multiple ideas at once. So I'll have multiple features and they're all on separate work trees that the agents are going to overstep on each other. And a perfect example here is I have a ghostly terminal running Herder. And I don't know if you can see it on the bottom left, I can zoom in. I have one, two, three, four, five tabs running Codex agents.