Member-only story
Your AI Model Isn’t the Product. The Harness Is.
What I’ve learned about harness engineering, and why I think it should be built once and reused across teams.
Someone on my team asked me a question recently that stuck with me. If every company can use the same models, why do some AI products feel genuinely useful while others feel like a smart intern who forgets everything by the next morning?
I’ve been chewing on that for a while, mostly by building things and watching where they break. The answer I keep coming back to is something people have started calling harness engineering. This post is my attempt to explain it in plain terms, along with an idea I think deserves more attention.
The model is only one part of the system
When most of us first used LLMs, the pattern was simple. You asked a question and got an answer. Now we expect a lot more. We want the model to draft a customer proposal using data that’s already in our CRM, or make a change across a codebase and run the tests, or pull the quarter’s numbers and write a summary for leadership.
Once you try to build something like that, you notice the model is doing less of the work than you’d expect. It decides what should happen next. Everything else, like giving it the right information, actually calling the systems, checking whether it’s allowed to do something, and keeping track of where the task is, has to be handled by…
