We’ve Been Building a Harness for Months and Didn’t Realize It
September 14, 2026 I went to bed at halftime of the Ohio State vs Texas game because it was way past my bed time and Ohio State looked dominant.
Imagine my joy when I woke up to see the news! Texas had handled the hard part, I just needed to figure out how to convert that schadenfreude into a newsletter about AI…so here we go.
An AI model is the talent on the field. Claude and ChatGPT are 5 star recruits making millions in NIL.
The harness is the coaching staff responsible for turning that talent into results. Ohio State fans can explain why the Ryan Day harness isn’t fully utilizing the talent on the field.
You're going to hear the word "harness" more often. It's the tech scaffolding surrounding an AI model that gives it access to information, tools, instructions, and a process for completing work.
The formula the engineers use is simple:
Model + Harness = Agent
I have no clue if this coaching analogy will hold up to the test of time, but I really just needed an excuse to hammer home the Ohio State collapse last night. Pure joy.
- The playbook: Instructions about how to approach the job.
- Play calling: Coordinating which tools to use and when.
- The scoreboard: Tracking what's finished and what remains.
- Film review: Checking the work and finding the mistakes.
The model supplies the intelligence. The harness puts it to work.
Coaching matters, right? James Franklin would blow away Ryan Day at wasting talent. Same for the harness. Somebody ran the same model through seven different harnesses on the same benchmark. Same brain, no changes. The scores ranged from 58% to 74.7%.
A 16 point spread, based entirely on the coaching/harness.
You can have exceptional talent and still struggle with execution. You can also improve execution without changing a single player.
That second point is the one worth sitting with, because when an AI tool disappoints us, the instinct is to blame the model. Maybe it needs to be smarter. Maybe we should switch providers. Maybe the next release fixes it.
Sometimes that's true. Sometimes we've handed a talented quarterback no playbook and a terrible offensive coordinator.
Anthropic documented a version of this with their coding agents. The models would attempt too much at once, lose track of progress, or declare victory without adequately testing their work. Adding structured steps, progress records, and explicit testing fixed it.
Two people on my team have been building harnesses for a while and I didn’t realize there was a word for it. Tom was hired as a junior relationship manager with zero technical background and now runs most of our automation through n8n. Ben spent two years getting our loan abstraction to 99% accuracy on the ugliest documents in this business, and almost none of that work was picking a smarter model. It was defining what a correct answer looks like and building the checks that catch a wrong one.
That's film review and coaching. Two years of it. That was transformational for us. I suspect you have something different but equally transformational for your organization. Invest your energy into building a harness to maximize the talent.
And if Ohio State fans think it’s rough being stuck with a Ryan Day harness, imagine being stuck with a Dabo harness. He’s still arguing why AI is a waste of time and money.