Project progress:
The machine works again.
As of today, the coding agent framework can handle small and medium-sized projects. It still fails on complex ones, which is inconvenient, because complex projects are the entire thesis.
Right now, the codebase is sitting at around 140,000 lines of code. I think I’m either at or near the limit of what the current generation of frontier models can handle before they quietly start melting the project.
Looking at it now, I can already see some of that happening.
There are too many compartmentalized, meticulously organized subsystems.
The thing’s a cathedral.
Reducing the cathedral
My honest feeling is that we can probably cut about two thirds of it without changing the process all that much. Now that I’ve figured out that we needed a custom ANN for the vector system, and that the agents need to reduce cleanly into swarms, the vision is much clearer.
Long-horizon solvers are tricky.
This isn’t the first one I’ve built, but it is the most complex process I’ve ever applied to one.
The hard part is that there still isn’t much public code for this kind of system. There are a couple of academic builds, a few credible demo apps here and there, but not many, if any, fully finished products that can complete whole projects at this scale.
Langchain is unintentionally interesting, and Hermes Agent comes really close at times, and then it fumbles the ball.
Without reliable references, you’re flying blind.
You are breaking new territory with a compass, a flashlight, and whatever survived the last experiment.
Honestly, that’s part of what makes the work interesting.
The process is:
Build.
Test.
Evaluate.
Rebuild and/or fix.
Audit.
Refactor.
Repeat.
The task running now, in a totally automated way, is the audit.
More soon 🕯️