FORGE — Self-Improving Software Workshop
An autonomous software-creation system that starts with trivial tasks, masters them through repeated build-measure-repair loops, then progressively takes on harder challenges while learning from every failure.
The Problem
I needed a workshop, not a one-shot agent. A loop that picks a task, builds it, validates it, critiques itself, records what went wrong, and gets harder over time — without me having to babysit it.
So I built FORGE.
The Loop
pick task → build it → validate it → critique it → record lessons → escalate difficulty
Each cycle adds to a lessons file that the next cycle consults. The difficulty knob steps up only when the previous level’s success rate clears a threshold I set per skill area.
What it’s built so far
628 runs to date — 348 passed validation, 280 failed and turned into lessons. Browse the full Forge Portfolio → for the names, scores, and domains of every run.
Why It’s Private
FORGE is internal infrastructure. Its lessons file is operational state from running my own systems — the kind of thing that’s interesting only to me but informative to a competitor.
The methodology (the loop shape, the difficulty-escalation rule, the lessons-as-input pattern) is in the /writing section. The source stays internal.