Heivol
← Back to Heivol

Articles

Notes from building Heivol — AI and ways of working, self-hosted infrastructure, and engineering leadership, plus a knowledge base on agentic engineering. Written by the people who run the stack, not a content team.

Essays

Longer pieces on strategy and ways of working.

Engineering · October 23, 2026 · 7 min read Nobody reads the assembly We stopped reading compiler output decades ago. If AI writes the code and nobody reads that either, the case for readable languages weakens, and the speed and energy we pay for them start to look expensive. Read the article → Personal · September 29, 2026 · 6 min read Australian again, with a knowledge base doing the legwork Born in Tasmania, Danish since 1995. How Claude and Contextbox carried an Australian citizenship case from a stalled 2023 attempt to approval: the wrong legal theory, the form, and the name problem. Read the article → Strategy · September 20, 2026 · 8 min read A web edition is not a strategy In the nineties every newspaper got a website. Schibsted also asked what the Internet did to the thing it sold, and today it owns no newspapers. Two questions to ask of any technology shift. Read the article → Ways of working · July 22, 2026 · revised Sep 17 · 7 min read Working with LLMs, as a daily craft Two years of daily LLM use, written up: a five-step loop where the model owns one step, why context decides quality, the gate nothing ships without, and the interest group that brings a team along. Read the article →

Knowledge base

Agentic engineering — a ten-part series on the layers that sit around a language model. Everyone benchmarks models; almost nobody benchmarks the thing around the model, which is where most of the variance lives. One layer per part, each with the failure modes, the tradeoffs and a diagram. Two of ten published so far.

Part 1 of 10 · September 22, 2026 · 7 min read Harness engineering A complete agent is five lines of Python. Everything between those five lines and something you would let near a production repo is the harness. What it contains, which half of it dies at the next model release, and the token tax it pays on every call. Read the article → Part 2 of 10 · September 22, 2026 · 7 min read Loop engineering Demos tell you nothing about iteration 40. The three ways an agent loop dies, why the model is a poor judge of its own completion, how to spot a thrash, and the difference between retry, replan and escalate. Read the article →