Reading

Adopted as a working method, then argued with: formalized mathematics still tames problems brute search cannot.

I studied how it found its audience as much as its content. By April 2026 I had decided the twelve-factor comparison was the wrong frame for my own work.

What I had been calling agentic system design under a better name. I adopted the term and published a design-space map against the post within the week.

It describes the pipeline I run over my own archive. Good confirmation.

Adopted directly as the grader loop in my skill tooling: a checklist, one change at a time, keep or revert.

The model for presenting a problem I had chased for two years, and for punctuating paragraphs with visuals.

It is easier to read than the original, and it is the model for the field-guide voice I want.

Reading these alongside the harness-engineering post is where my narrative converged on a Unix for agentic systems.

A primitive that deserves first-class support. It excels when you do not care when the work finishes, only that it can be checked. My system's day-and-night cycle is built on it.

It turns an agent session into a computer with a small surface area. The idea is bigger than the implementation, which is what made me want to build on it.

A kernel-mode form of self-improvement that no lab would ship. It cleared away my search for advanced agent use cases, and sixteen of its agents now run my businesses.

I agreed with all of it and realized I had been doing it in private. Writing as preparation, with the scripts and logs as the work.

Six essays in one sitting. The one on cities gave me the idea that environment is a force you use rather than fight, which became everything I later called prevention architecture.

The justification for building a language from the bottom up while nobody yet knew how to program with AI.

I re-read it every few years, which is its own kind of data.

I subscribed in 2024. By 2026 what I most wanted to learn from it was how it reaches people.

Its list of strategic behaviors is a persistent mode of being an agent should embody, not a skill you summon once. Adopted into an agent design.

It has minimal fluff and no fake anecdotes, and it taught me a test: if you cannot name the specific thing, it is hallucinated.

Composable prompts from the command line, and the command-line shape of my own tools starts here.

A Python-to-Rust smart-contract compiler that was one of my own early ideas, shipped by someone else. It got me to ship something within the month.

An insight I had had a year earlier without externalizing it. The question it left me with: what is lost at the filesystem level that the interpreter level keeps?

It is my example of a dependency you can demand-page rather than re-derive.

Defining models with code, in 2024; I never got past the intent to study it properly.

It is about the time-versus-compute tradeoff for humans and machines.

It was the learning path that made sense to me in 2024.

One thing stuck: training on code improved reasoning, as if language itself is the space where learning happens.