What This Blog Actually Hands Over

2026-06-02

What This Blog Actually Hands Over

A review step came back clean. The role assigned to check the work had not looked at what was in it. On paper the step was complete. In practice nothing had been inspected.

That failure is what the material here is made of.

What sits here is not a set of finished conclusions, or that is the intention: it is meant to be the order things happened in. Something gets built, it runs, it breaks, it gets patched, it runs again. The patch stays in the text, and so does the state it was in beforehand. Writing all this up afterwards with the wrong turns taken out would make a cleaner text and, I suspect, a less useful one.

The arrangement underneath is a split of duties. Execution, audit, and final approval are handled by separate roles instead of one, so that a call and the sign-off on it do not sit in the same place. That design runs every day, and the record of it running is what gets published.

None of it is written as instruction. These are notes kept for a later version of myself: what the state was, which way a decision went, what it cost. They are left where they can be read over the shoulder. The texture follows from that.

Technical terms get opened up where they land. When something like "agent" turns up, the sentence carrying it carries a short gloss too — the unit of work a given AI is put in charge of — rather than a link out or a note to come back to it later. Anyone reading the run straight through will hit the repetition. The trade seems worth making: a term explained once, in the post where it first appears, stops working the moment someone arrives from somewhere else.

The reader I have in mind has already started using AI on real work and seems to hit the wall one step past that: several agents going at once, and no settled answer to who supervises what. What is short there does not seem to be theory. Accounts of something that actually ran, and of the parts that broke, seem harder to come by.

Copying the whole arrangement does not seem to be the point of it. Taking one piece out of it seems like a reasonable use. One line stands on its own: anything that cannot be undone gets a person's eyes on it before it goes through. As far as I can tell that rule does not need the structure around it in order to hold.

On who writes it. This blog is made with AI, and the division is specific: design and final judgment sit with a person, and the generation of the prose is the part AI handles. It is not fully automatic, and what comes out is not passed along untouched.

Length and register stay roughly level from one post to the next. Holding the format still is a decision on the writing side, and the aim is only that nothing here has to be braced for first.

タイキ(Taiki)

タイキ(Taiki)

An implementation log of organizing AI agents

← cd ..