i think i'm a mad scientist

I can code (some). It will work (sometimes). The quality of that code might keep some experienced software engineers awake at night though.

The thing is I like to make things.

The problem is I don’t always have time or capacity to devote to making those things. There are often weeks or months between my working on my projects. This era of agentic development has fundamentally changed my capacity for homelab shenanigans, and I think my ability to forget what my own code does, might have accidentally turned out to be an extremely useful qualification for building with AI agents.

You see, we share the same failure mode, agents and I. Neither one of us remembers what we did last time.

I rarely came back to one of my projects remembering why I’d made the decisions I’d made, how things fit together, and sometimes it was as bad as, “what was I even trying to accomplish here?” As a result of that, when I started little projects like my agent doorbell or bigger ones like my wife’s pilates streaming platform - I designed the workflow not just around build this thing, but build this thing I’m not sure that I’ll come back to any time soon, so make sure everything is predictably organised, clean & structured.

So everything that I could mechanically enforce, I enforced. Because my discipline is unreliable, I expect the agent’s discipline to be unreliable - but I can give it something I never had - enforced consistency. So it was with mad, incredible, and maniacal power that I created structural blocks for the agent that would prevent progress until the conditions were satisfied. This doesn’t just apply to simple objective metrics either, subjective stuff gets evaluated and flagged too — the agent raises it, I rule on it, and the ruling gets captured.

Everything is captured.

It doesn’t matter if I leave in the middle of building, come back with a small patch, a meaty bug, or a whole new feature - it all routes through the same mechanisms, the codebase, documentation, agent & human context are all kept in lockstep.

I write all of this because there’s a bit of a debate running right now about “why software factories” fail - whether agents are incapable of maintainable code, or whether we’ve just put them in environments that don’t enforce it.

I can’t settle that from a homelab. Not that anyone is expecting me to… What I can tell you, is that in my personal experience, considering my own shortcomings when designing the framework my agents work in, means that I have confidence that I can always pick up where I left off smoothly - and things never end up in a state of pure chaos… at least for now.