What you can read

What you can read here is the real, ongoing process of matsumotory running his own research and development together with AI. AI agents write code and prose, another AI does a first-pass review, and in the end he decides what to merge. He runs this loop every day and writes down what worked and what did not, exactly as it happened. It is not a story cleaned up after the fact; it is the raw operation of a workshop that is running right now.

Practical knowledge, while it is fresh

At the center is practical knowledge about how to design an operation that hands development to AI agents. For example: when humans can no longer keep up with reviewing the changes AI produces in volume, where do you start so the flow works again? When you let AI judge the quality of code and prose and the verdict wavers every run, how do you make it steady? How do you separate workspaces so several AIs can run at once without breaking each other's work? Each is written in the order he actually hit the problem and fixed it. The bar for publishing is simple: if you strip out the measured numbers, the story of the failure, and the reason a decision was made, nothing should be left.

The other thread is the knowledge from the frontier fields he is tracking in his research and development, readable while he is still in the middle of tracking them. He writes about new tools and methods from the vantage of an infrastructure researcher, not as marketing or a tutorial, but as the result of trying them with his own hands. The center is technical, but there are also essays on what he thinks around the edges of development, and entries on the joys and pains of building together with AI.

With AI, both the volume and the way of working changed

This operation is not an idealized picture; both the volume and the way matsumotory works actually changed. Volume first. As a yardstick anyone can check, compare the yearly contributions on his GitHub profile. In his busiest three pre-AI years, 2013 was 1,071, 2014 was 1,878, and 2015 was 1,674, for a three-year total of 4,623. In 2026, the year he started working with AI in earnest, the count up to the data cut-off on July 17 is 6,183. In about half a year, that passed his busiest three pre-AI years combined. Divided by days, that is about 31 a day, and the squares on the graph are filled almost every day. It is a volume that runs as a daily habit, not a one-time burst.

The interesting part is that not only the volume but the way of working changed. Normally, to run up numbers it is faster to pile commits straight from your machine, yet the share of pull requests in the contributions rose from a few percent in the pre-AI years (3 percent in 2013, 1 percent in 2014, 7 percent in 2015) to about 30 percent in 2026. That is because he came to run, at high frequency, the professional pattern of branching per feature, turning it into a pull request, and merging only after an AI review. It shows up in that share: the volume went up while the discipline held, rather than getting sloppy for speed.

The scope widened too, adding more system-level tooling on top of web application development and widening the languages involved. Counting commits alone, they went from 128 in his busiest pre-AI month to a monthly average of 1,227 from March 2026 through July 17, about 9.6 times as many. Month by month, June was 1,418 and July reached 1,726 in just 17 days, again rising toward the present.

So the numbers do not walk off on their own, here is the honest caveat. July is a partial count up to the 17th, and counting it as a month makes the multiple look conservative. These are actual figures counted from matsumotory's public activity on GitHub. What matters is not any single multiple but that you can read, as a firsthand record, the actual way of working in which volume, process, and scope all changed together, and what happened along the way.

Who this is for

It should be most useful to engineers building alongside AI in real projects, to people who want to learn the operation of agentic development from concrete examples, and to those who want to know the frontier of research and development early. There is more on fit and mismatch on Who this is for.

About the author

This is written by matsumotory, a researcher in internet infrastructure who has long kept up both research and open-source work. His background, and why he is trying to preserve himself as an AI, are on About matsumotory. For how the whole site works and the thinking behind it, see About.

The free articles, and the summaries of every paid article, are made to be worth reading on their own. Sample articles, the full text of pieces that are otherwise paid or subscriber-only, are also open for free, one per plan. If you want to go deeper, dig into knowledge with Matsumotory-kun, or support the project, subscribing adds a fuller experience. The plans and the samples are on Plans.