I've run into a really interesting failure state.
See, it turns out your own documentation can rise from the dead and bite you.
Let me explain.
I've got two pieces of the same system that I'm building. Both agent systems.
The first is Pillowfort, which is basically an orchestrated theater company of agents, tuned to build and run real-life businesses.
The second is Qwench.
Qwench is a sandbox where I can do two things:
1. Determine whether I can turn Qwen3.5 9B into a capable local-only coding agent that runs on a five-year-old gaming PC, without losing the process discipline or capabilities you'd expect from a CLI coding agent.
I think I can.
Pretty sure, actually. It's a model under 50B that can reason about whether it's being evaluated. That’s a promising starting point for turning it into a coder.
2. Experiment with features and workflows that I can't safely test in the larger system.
Pillowfort isn't an experimental build anymore. It's a product. Something normal people should eventually be able to use.
Qwench lets me bridge that gap. It gives me space to fail and develop new components without having to manage a self-generating system, rules engines, security layers, context management quirks, and all the bureaucratic guardrails I wrote to keep the larger Gödel machine functional.
But the development process revealed something interesting.
Both projects draw from two years of research and hundreds of thousands of lines of code. Somewhere along the way I started reusing older repos instead of rebuilding everything from scratch every iteration.
That was Sam's idea. He's my business manager and spiritual soul sibling in these matters.
The problem is: I document everything.
Aggressively.
And documentation is infrastructure.
Usually.
Except when you're using your older repositories as spare parts.
That's when things get tricky.
Because if you're not careful, documentation from unrelated projects... where you're only borrowing a piece here or a component there -- can derail your coding agents.
When you're using AI code generators to build AI code generators, documentation stops being passive text in your repo.
It becomes part of the instruction surface.
Your agents read it.
They treat it as architecture.
They follow it, even when it isn't on the roadmap anymore.
In other words:
Old documentation doesn't just sit there.
Rather, it might start steering the system.
Good to know.