Shipping

What this series is about ​

Generation moved first and nothing downstream moved with it. That is the whole subject.

The posts follow one piece of work through the parts that did not get faster. It arrives at review, where the queue is now the place code gets finished rather than checked. It gets merged and needs to be observable, because the usual argument for skipping instrumentation was that you understood the code, and that argument is weaker when nobody typed it. Then eventually it pages somebody at an hour when the person holding the pager did not write a line of it and cannot ask the author anything.

Each of those was a solved problem when a human wrote the code and held the context. None of them is solved when the context never existed outside a session that has since been cleared.

There is a measurement running through the series that I keep coming back to. The numbers everyone quotes about AI and delivery are almost always about how much code arrives, and the interesting ones are about what happens to it afterwards: bugs per developer, incident-to-PR ratio, how much of the review queue is spent bringing code up to a standard it should have met before the pull request opened.

This series is short and specific. It is not a devops guide, and it assumes you already know how to run software. The question is narrower: what changes about that job when the author is a model.

Where this is going ​

The gap is the deploy step, which currently sits between two posts that both point at it. Feature flags, canaries and the rollback that a database will not honour are the piece I want to write next, and it needs a real migration going wrong rather than a diagram.