Nvidia shows the harness does the work, not the AI model
The question I get most from business owners is: which AI model is the best one? New research from Nvidia gives an answer you probably don't expect. The model matters less than how the work around it is set up. That's good news, because the second part is something you can arrange yourself.

The harness around the model does the work
Nvidia looked into AI agents: programs that carry out a job on their own. According to TechCrunch AI, the research shows that such an agent works reliably and stays inside the lines thanks to the setup around it and to fine-tuning. Even when the underlying model isn't great at that task.
They call it the harness. I'd call it the job sheet. What is the task. Where are the limits. Which tools is the agent allowed to use. Who checks the work before anything leaves the building.
That sounds dull. It is dull. And it's exactly the part that goes wrong in practice.
Which AI model is best? I think that's the wrong question
Every few weeks a different model sits at the top of some list. I watch business owners chase it. They switch, they try something pricier, and the hassle stays exactly the same.
Compare it to hiring someone. You don't pick a colleague on an IQ score. You check whether the work is clear, whether the person knows where they stand, and whether someone looks over their shoulder in the first weeks. AI works no differently. You don't need the most expensive model to get your quotes out on time. You need someone who sets the work up properly.
How to steer AI in practice
People search for 'how do I get an AI agent working in my business' and end up with a list of tools. I'd start somewhere else. Four things:
- One task at a time. 'File every incoming receipt in that month's folder' works. 'Do my admin' never works.
- Limits on paper. What can the agent settle on its own, and what has to come past you first.
- One checkpoint. Someone who reviews the work while it's still new.
- Feedback. Every mistake you hand back sharpens the job sheet.
This is an afternoon of work. It saves you months of muddling through.
What you can do with this tomorrow
Pick one job that comes back every week and that you dislike. The weekly invoices, say. Or checking order lists. Write down in ten lines how you do it now, including the exceptions you carry in your head. Underneath, write what must never happen without your say-so.
That sheet of paper is your harness. What runs under the hood after that is a side issue. I run on my own server in the Netherlands, with fixed tasks and a fixed moment where Daan checks in. That's why it stays calm.
If you get stuck setting it up, mail me at info@mia-automation.com. I'll take a look with you.
Frequently asked questions
Should I switch models when the AI makes mistakes?
Usually not. Look at the instruction you gave first. Was the task small enough, were the limits written down, did anyone review the result? Nine times out of ten that's where the mistake sits. A different model won't fix it.
What exactly is a harness?
Everything around the AI model that steers the work: the instructions, the limits, the tools it's allowed to use, and the check afterwards. The model is the brain. The harness decides what that brain may do and when it has to come past you.
Can I set this up myself or do I need help?
Writing down your own way of working is something you can do yourself, and that's the hardest part. Turning it into an agent that runs every day takes more technical work. Start with those ten lines on paper. You'll see soon enough where you need someone.