How Long Does Harness Engineering Have?
Everything useful in a harness ends up absorbed into vanilla. What survives is what the vendor structurally will not or cannot build. Notes on the shift.
Everything useful you can build into a harness eventually gets pulled into vanilla. What is left after that is not the useful features. It is only what Claude structurally won’t or can’t do. The harness-maker I spent a year on taught me that. When Claude Code shipped Goal and Dynamic Workflows by default, it covered the exact spot where harness-maker’s loop and review had been.
It started simply. Harness engineering was having its moment, and I wanted a harness of my own. So I built harness-maker and put a full year of experience into it. Every project has its own taste, and I needed a way to work that taste into the harness. Between company work, side projects, and things I actually sell, no single open-source harness could fit all of them. A project that runs Spec Driven and a project that runs light on Task Driven need different workflows from the start.
Then I saw Dynamic Workflows and changed my mind. Most OSS harnesses, I thought, were about to be absorbed. My harness-maker’s loop and review would be no exception. It made more sense to count them as already absorbed.
So I find myself asking what will be left of harness engineering. We moved from prompt to context, then from context to harness. Will this one also become a term from a past season, the way prompt engineering and context engineering did?
Lately I notice more people around me going back to vanilla Claude Code. I understand why. Vanilla will keep getting stronger and more personal, and anything worth using will move into native. Then, by definition, only one thing is left. Once the useful features have all been absorbed, what remains in that spot is what Claude structurally won’t or can’t do.
Vendor-neutral features are one example. Vendor-hostile ones are another. Those survive. A tool like ccusage is one of them: it tracks your token cost from local logs and helps you spend less. No vendor is going to build something that cuts into its own revenue. Methodology taste will hold on for a while too, I think. By methodology taste I mean the workflows some OSS harnesses have encoded, such as a workflow that forces SDD, or a method that forces a deep interview when you plan or define a spec. Things like that.
That does not last either. The moment a dynamic taste setting or a methodology setting proves useful, the vendor puts it into vanilla by default. Then it is absorbed.
So I have stopped piling more harness onto spots that are clearly headed for absorption. I spend my time only where the vendor structurally can’t or won’t go.
Looking back, OSS and personal harnesses alike are a training signal that teaches the vendor my taste. The better I encode it, the faster the vendor learns it, and the faster the harness goes stale. For that short transition window, though, it is genuinely worth having.
Comments
Loading comments...