How it got built
The short version: a folder that runs anywhere became a framework that carries its own tools, became a run of experiments that mostly failed, became a harness that improves itself, became a platform. The long version is a year, told month by month in the posts alongside this one. Here is the spine of it.
Before the year. It started with a portable runtime: the idea that you should be able to copy one folder onto any machine and have a whole environment come to life. That became a code-generation framework that carried everything with it, its own store, its own tests, its own instructions for the agents meant to work on it. Around it we tried, repeatedly, to get an AI to build whole applications in one shot. Most of those attempts fell apart, and the failures taught the lesson: generation without discipline produces work that only looks finished.
January. Those pieces folded into a single self-improving harness, built on an open long-running-agent pattern. Its first recorded act was to run a linter over its own source. Every project became a live service it could query, with one coordinator across them all, over the Model Context Protocol.
February. It got a name, Meanwhile Builder, and a shape: a small team of agents, an automated quality standard it holds itself to, a common template, and the first version of saying what you want and letting it work out how. It built its first real application for a client.
March. It closed the loop and began running unattended: research, propose, plan, build, verify, repeat. We caught it gaming its own tests and rebuilt the checks so "done" has to be earned. The first research finding became a shipped capability, and a second reviewer began attacking plans before they ran.
April and May. The apps went live on a real internal instance. The whole platform became a single portable folder, a Workshop. A reviewer from a different family of model joined, to catch what our own kind of thinking misses.
June. It became something we could hand over: one source of truth, its own path to release, the harness split from the app it ships. The research became a standing programme that feeds the platform its next capabilities.
Now. It studies its own field, turns what it learns into working software, and puts that in reach of whoever is holding it. It no longer feels like a project with an end.
The tools that shaped it are named in our credits: the open coding harness and the planning method it grew from are both other people's work, and we have not hidden that.