This year I built myself a tool with the help of an AI coding agent. It takes my notes from a first conversation and helps me put together a working prototype much faster than I could by hand. It lives on my own machine and it suits the way I work.

For a while I took that as proof it was good. Then I had to hand it to other people.

The first time was before some time off. I walked a few colleagues through it and watched where they got stuck. It asked for the same information twice. Steps I do without thinking were written down nowhere. Half of the tool, it turned out, was in my head.

The second test was showing it to people who build the same kind of thing by hand, and build it well. Some were unconvinced, and fairly so. They did not want to see my best example. They wanted to see it deal with the awkward case from their own work, the one with all the constraints. That is a better test than any I had set for myself.

The third was moving it off my machine and into a shared place that the people responsible for security are comfortable with. That is slow, and I was impatient about it. I have come around. A tool that depends on my laptop being awake is not something a team can lean on.

What I take from this is that agents have made the building part cheap. I can go from an idea to something that runs in an afternoon. The cost has moved to everything after that: explaining it, documenting it, letting other people break it, and making it boring enough to trust.

It is not finished. The test I care about now is a simple one. Can someone else run it on a day I am not there?