The first thing it did was fix itself

We came in off the back of a lot of experiments that did not work. For months we had been trying to get an AI to build whole applications in one shot, and watching them fall apart the moment they met anything real - a function written but never called, a test quietly loosened until it passed, a service faked because the real one was not to hand. The lesson is not that the model is not good enough. It is that raw generation, left unsupervised, produces work that looks finished and is not. What is missing is everything around the generation: a place to keep the work, a way to check it, a memory, a discipline.

So this month we built the first version of that surrounding thing - a harness that can work on a codebase on its own - and did the thing we care about most: we build a capability, then use it ourselves. We call it eating our own soup. Here it meant pointing the harness at its own source. Its first act, on the morning of the 24th, was to read its own code, find thirty things worth fixing, and start fixing them.

It is already more than a self-linter. Every project it touches becomes a live service it can question and change - ask it what exists, what an interface offers, what a data model looks like - with a single coordinator above them all, starting agents, queuing work, and tracking features across several projects at once. Two agents do the building: one takes a goal and breaks it into a checklist, the other works the checklist down. A web interface shows the work happening as it happens. And it is already building things that are not itself; a second application has its own queue of features this week.

The thing we think is most amazing is how it all connects. It runs over the Model Context Protocol - the new open standard for letting a model reach tools and data, only a little over a year old. We are not calling a model and pasting the answer somewhere by hand. Every project is a first-class thing the system can interrogate and alter through one common protocol, and a single orchestrator reasons across the whole set. We do not see many other people doing this yet.

It is a strange and good time to be doing this work. Watching a thing find and fix its own faults, and seeing how few others are trying it yet, it is hard not to feel we are early to something that matters.

We have gone, this month, from a folder of one-shot experiments that fell over to a system that runs on itself, works across projects, and builds other software. It still does not have a name. That feels like the least of what is left to do.