Pointing it at itself
This month it took a name and a shape. We called it Meanwhile Builder - Builder, for short. Not a clever name, just what it is.
We stopped treating it as one loop and started treating it as a small team. There is a base agent now, and specialised ones on top: one that writes features, one that reads code for quality problems, one that fixes them, with a thin coordinator handing work between them. The same harness that could improve its own source in January can now run a whole crew at a task.
The part we are most pleased with is that it holds itself to a standard. We wrote out eight rules the platform's code has to meet - the shape of a project, how things are named, where logic is allowed to live - and taught it to check every project against them and fix the small violations on its own. It is the first time the quality of the work does not depend on someone remembering to look.
We kept eating our own soup. It went back over its own earlier work and wrote the research it should have done the first time, and it now starts every new project from a common template, so they all begin the same clean way.
Something else started this month that we think matters more than any single feature: a change in how we ask for things. Instead of telling Builder how to do a job, we began telling it what we want and leaving the how to it. Right now that means writing a plan in plain language and letting it turn the plan into work, build it, and check the result against what we asked for - build, check, fix, with no person in the loop for the ordinary cases. Where we are heading is further: to say what we want, have it research the established way to do that thing, the best practice and the industry-standard method, formalise that into a process, follow the process, and hand back the result. We are at the start of that, not the end of it. But the shift from telling it how to telling it what is already changing how the work feels.
At the end of the month it built its first real application for someone outside these walls - a themed workspace for a client. Not a toy, and not itself: a thing a customer will use.
We have gone from an unnamed self-improving script to a small team of agents that enforces its own quality bar, starts from a consistent template, executes plans end to end, and has shipped its first real app. It is starting to feel less like an experiment and more like a way of working.