← Posts

Procrastination with excellent tooling

Most agent setup work is procrastination with excellent tooling. The tell is that the builder cannot name a single shipped outcome it caused.

4 min read

Harness work can be useful. It is also my favorite form of procrastination.

The model is a component. The harness is everything around it: what context it sees, which tools it can reach, what it is allowed to write, how failure is caught, what happens on retry. Some of that work changes what you can ship. A lot of it is engineering-shaped activity with no natural completion condition, and other engineers recognize it as progress. That is what makes it such a good way to avoid the thing you were supposed to ship.

So I need a diagnostic I can apply to myself. Name one shipped outcome the harness caused, or one failure it caught that would otherwise have escaped. A feature list does not count. A cleaner abstraction does not count. If I cannot name the outcome or the catch, I was tooling.

Applied to my own work, at the level I can stand behind: ContextStack is a cost-aware multi-agent coding harness. The architecture I will name is a Laravel orchestrator, a Node interface, a universal agent harness, and a model router. I have built four iterations of that stack. ContextStack.com publishes the architecture. The implementation is private, and the site is in development, not a product I am handing you.

Four iterations is already a smell. Iteration is the unit harness work prefers, because each one feels like a decision and none of them has to face a user. I can describe the layers. I cannot tell you, version by version, what shipped because of them, and I am not going to invent a release history so the confession has chapters.

OpenClaw was a side route I played with that did not become the path. I also downloaded Hermes and never used it. That is the tell in miniature: downloading a harness feels like research, and not using it is the result. The unused download is the most honest artifact I have of this habit.

Why the work is such good procrastination is not mysterious. It is engineering. You can spend a day on routing, retries, and permissions and have a diff that looks like senior work. It has no natural end: there is always another model to wire, another tool to wrap, another rule that will make the next run safer. And it is legible. A person who spent the afternoon on a harness can describe the afternoon. A person who spent the afternoon not starting the product cannot.

The diagnostic is supposed to cut that, and applied retroactively, most of my iteration history fails it on the evidence I am willing to publish. Dark Factory's news pipeline is a shipped publishing system, and editions go out. That is a real outcome of the broader agent estate, the scheduled research and news work I actually run. It is not a receipt I can pin to iteration two, or three, or four of ContextStack, and I will not paper over the gap by assigning it one. I am also not claiming the pipeline runs itself. A publishing system that exists is enough. A story about lights-out authorship would be a different essay, and I do not have the receipts for that story.

What survives the test is not a version number. It is the standard. If I cannot point at a ship or a catch, the harness work was a hobby with excellent taste. The public architecture can still be worth writing down. Writing it down is not the same as the architecture having earned another iteration.

A boring harness, as I want one, is the opposite of a fourth rewrite. It is the smallest set of parts that changed a ship or caught a miss, with everything else removable, including the parts I am fond of. I have not done that removal in public. Which of the four iterations would survive the test without mercy, I do not know. The unused download would not. The side route would not. The rest I would have to judge honestly, and the only way I know to think that through is the one thing I should not do, which is write a fifth iteration to help me think.