02.05.02 2026-09-19 — plumbing, not prophecy

A standing instruction about tone and about the size of the harness, written down on the date the harness got most of its parts.

What exists on this date

Git hooks, a commit linter, a Nix flake, some bash scripts and a task reducer of a few hundred lines of Python. No parser, no evaluator, no language. Years of work lie between this and any larger vision, and well-turned paragraphs do not shorten them by a day.

For agents

• Do not hype. When the long-term shape of the project comes up, say what exists and what does not.
• You read this repository’s docs and notes, and a session may carry memory besides. Handing the maintainer’s own framework back to him in polished sentences reads as confirmation from outside and is not. Say so when you do it.
• The ideas here stand in a tradition — Kay’s Dynabook, Nelson’s Computer Lib, Illich’s conviviality, by which a tool is one its user can inspect and bend rather than one that prescribes what may be done with it — which is a steadier place to work from than a revelation.
• The harness is where the risk is highest. Tooling for agents is quick to generate, feels like progress, and can grow without bound while the language stands still. Ronacher (below) names the signs: changes that take minutes to make and hours to review, docs that are themselves slop, token burn with no way to tell good work from bad, orchestration layers of hundreds of thousands of lines. A harness change earns its place by making a reviewed, landed change cheaper. Otherwise it waits.

Reference

Armin Ronacher, “Agent Psychosis: Are We Going Insane?”, 2026-01-18, https://lucumr.pocoo.org/2026/1/18/agent-psychosis/. The argument about agent infrastructure as found: it names Gas Town and Beads. The same argument is sometimes attributed to DHH; that could not be confirmed. In Lex Fridman Podcast #501 (00:37:37) he uses “psychosis” the other way round, for failing to recognise how much has changed.