In the ten days to 2026-10-12 we shipped 156 changes across four repositories. 100 of them went into Orqestra, our agent system, and 48 into a game. The rest were spread over our website and a fitness project.
Most of that is not interesting to read about. A few pieces of it are, because they show what an agent system is actually made of when you open the panel and look: a set of small, specific decisions about who gets to do what, and what the screen should say when something goes wrong.
This post is for people who do not write code but are starting to run work through AI. We use our own week as the example, and we stay with what the record says.
What 156 changes in 8 days means
Across the window, changes landed on 8 different days. Together they touched 731 files, adding 26424 lines and removing 6025. Per repository: 100 commits in Orqestra, 48 in skat-game, 6 in NeuraGrowthHTML and 2 in FitGro.
The first of those changes landed on 2026-10-05 and the last on 2026-10-12. If you picture an AI workflow as a single clever prompt, that is the wrong picture.
One caution about the numbers. A big count of lines changed says how much was touched. It does not say how much was gained, and we are not claiming that it does.
The panel has to say what the system does
On 2026-10-08 we shipped a package from an audit, called Audit 08.10, package D. The idea in its title is the one we would keep if we forgot everything else: the panel says what the system does. It touched 40 files, adding 741 lines and removing 147.
It is a list of small honesty fixes. A failed book can now be resumed instead of the regenerate button refusing it. A run that is still going shows what its calls have spent so far. A bundle's price on the matrix counts its own parts. A backend that does not answer now shows "serwer nie odpowiada", which is Polish for the server is not responding, instead of sending you round a login loop.
A tool that fails silently, or fails in a way that looks like something else, costs you more than one that fails loudly. A login loop that is really a dead server wastes your time, and you cannot tell why.
A step where a person still decides: the Etsy review
Another item from the same package: the Etsy review modal now starts on the product's own tier. That is a small sentence, but it points at a design choice.
We cannot tell you from the record how the review decision is made beyond that, so we will not invent a process.
We have written before about what happens when Etsy listings expire and drop out of our catalogue. That is a different failure, and a reminder that the Etsy side of a catalogue is something we keep having to watch: https://neuragrowth.co/log/an-etsy-listing-expired-and-dropped-out-of-our-catalogue/
Turning a series into one guide
On 2026-10-07 we shipped a change called Renewals: a series merged into one premium guide. It touched 13 files, adding 1681 lines and removing 27. The reason is recorded in the commit: the founder asked, on 07.10.2026, for a trilogy (products #111 to #113) and its bundle (#114) to be renewed as one comprehensive premium guide on current facts.
The record is clear about the gap. No existing path did "from N products, one". The rewrite keeps a part's place in its series, the bundle rebuild only merges PDFs, Regenerate stops at the old proposal's cap, and the retire step knows no successor. Five series with bundles stand in the catalogue, and three of them have expired parts.
So we added a button, labelled "Scal w jeden poradnik premium" (merge into one premium guide), on a bundle's or a part's renewal card. It runs as a background task. We wrote it up in more detail here: https://neuragrowth.co/log/celery-task-merges-a-product-series-into-one-guide-13-files/
What the alerts looked like
Our children's series runs on a schedule, one new episode in a slot each day, produced overnight by an automatic run. On 2026-10-11 at 18:00 an alert told us the buffer for the counting series was empty and that the night run at 03:00 would try to make the next episode. If that failed, the slot would go out empty for the day. The alert named the cost of having no reserve: a series with no stock depends on a single night.
The same alert listed the next seven slots, and in the record several of them were marked held, waiting for a decision. On 2026-10-12 at 01:00 another alert reported that the topic queue for the counting series was empty and had been refilled automatically with 9 topics from the engine's cycle. A third, at 14:00, warned that the 14 October 16:00 slot had no new episode and that a reserve episode would go out if the night did not make one.
The record does not say why the nights fell behind. The alert itself says only that a gap in the next two days means the nights are failing or not keeping up, and points to the run list. We will leave it there rather than guess a cause.
What we would do again
The buffer alert did not say something is wrong. It said which episode the night would try to make, and what happens if it fails.
If you are choosing tools for the image side of a pipeline like this, our comparison of three image models is here: https://neuragrowth.co/blog/stable-image-core-vs-ultra-vs-flux-schnell-which-for-filler/
If you run work through an AI system, do not judge it by how clever the model is. Judge it by what the screen says when something goes wrong, whether a human review step opens where the reviewer needs it, and whether an alert tells you what happens next if you do nothing.
Originally published at neuragrowth.co. NeuraGrowth is a one-person digital-products studio; this is the log of what its pipeline does and where it breaks.
Top comments (0)