Fifty years of things that had to stay up.

I wrote my first program in November 1975, on an IBM 360 at Dartmouth’s Kiewit Computing Center. Since then I have worked on systems where being down was not an abstraction — including the NASDAQ trading engine, where the cost of a bad deploy was measured in ways that focus the mind.

I now build AI-integrated applications, and I keep noticing that the hard parts have not moved. The vocabulary changed. The failure modes did not. Retry semantics, idempotency, blast radius, the discipline of knowing what your system does when its dependency is slow instead of down — all of it transferred intact, and most of the current writing about agents skips it.

That gap is what this publication is for. Not tutorials; a model can write those. Judgment under production constraints, with the reasoning shown.

What you can expect

One free essay a week. One deep dive a month for members, with a repository you can run and a decision record that explains what I rejected and why. When I am uncertain, I will say so. When I am wrong, I will say that too, in the same place I was wrong.

Elsewhere

I write here as an individual. Load-Bearing Code is published by Aegis Tensor, my own company; nothing here represents any employer, and I do not write about client or employer systems.