https://piotrgoral.github.io/Recent content on AI from first principlesHugoen-usSun, 13 Sep 2026 00:00:00 +0000https://piotrgoral.github.io/posts/what-is-a-harness/Sun, 13 Sep 2026 00:00:00 +0000https://piotrgoral.github.io/posts/what-is-a-harness/<h2 id="in-short">In short</h2> <p>I hear the word &ldquo;harness&rdquo; several times a day, and it seems to mean something different each time. After a while, it gets overwhelming. What belongs where? How do you compare harnesses? And where do you start if you want to build one?</p> <p>So I went through research papers, blog posts, books, and podcasts to understand how each source defines a harness (chapter 2). Then I proposed my own framework: prompt, context, LLM agent, and a deliberately narrow harness core: environment, turn lifecycle, and verification (chapter 3). Once that core is in place, the framework can be extended with cross-session memory, interfaces, and observability (chapter 4).</p>