Software that builds itself.

LEVEL 1 · THE PRODUCT LOOP

It finds and ships its own features

1. 01

   Listen. Feedback from the support team (“no, we don’t answer it like that, because…”), my feature requests, anything odd in the production logs.

2. 02

   Propose. Whatever matters becomes a change request against itself — and an entry on its own roadmap.

3. 03

   Defend. A request grows into a full spec only after surviving cross-examination by GPT and Grok.

4. 04

   Ship. A long pipeline in which every quality gate has to pass: tests, independent reviews, security, docs.

LEVEL 2 · THE PROCESS LOOP

It improves the way it builds itself

After every run, an LLM council reviews the machine, not the product: where the cycle lost time, what slipped past a gate, which rule would have prevented it. Then it rewrites its own harness — the skills and rules its agents work by.

• Every process change is a hypothesis, tracked over the next runs until it is confirmed — or refuted and reverted.
• At most two small, evidence-backed changes per run. Deleting a rule beats adding one.
• No evidence, no finding.

Control engineers call this a self-tuning regulator: one loop runs the plant, a second loop re-tunes the first. I’ve spent twenty years building the first kind. This one has both.