The AI Model Is Not the Product. The Harness Is.

The Short Version

• The AI model is the active ingredient. The harness, the system of checks, tools, memory, and workflow around it, is what turns that ingredient into a working agent.
• One 2026 benchmark swapped only the harness, same model underneath, and task performance moved about 24 points. The scaffolding is not a detail. It is most of the outcome.
• When you evaluate an AI tool for your practice, the model name tells you maybe half the story. Judge the system, not the model.

A late-night conversation in Houston

The AI model is not the product. It is the active ingredient.

I was at the Texas Pharmacy Association annual conference in Houston this weekend, and one of the best conversations I had was a late night with my longtime friend and mentor, Raj Chhadua, about models versus harnesses. It is a distinction most people in healthcare have not caught yet, and it matters more than the brand name on the slide.

Here is the confusion I still see. People assume that whether an AI tool runs on Claude or OpenAI on the back end, that is the whole story. It is not. The model is the active ingredient. What turns it into an actual agent is the harness, the tool calling and the system wrapped around it.

I have been building AI models and agents for the better part of three years. So when Nimbus Healthcare was accepted into the NVIDIA Inception program last week, it did more than put us alongside the teams building the next generation of AI infrastructure. It sent me down the path of NVIDIA's NeMo work and their newest framework, NOOA.

The pharmacy version of this idea

I am a pharmacist. Most of my colleagues in healthcare are clinicians, not engineers. So when a technical paper drops that actually matters for how AI shows up in our clinics and pharmacies, I want to translate it. This is one of those.

NVIDIA Labs published NOOA last week, along with a paper on what they call "harness engineering." The term sounds obscure. The idea is not.

Here is the plain version. An AI model is like a brilliant clinician with no chart, no workflow, and no verification steps. The harness is everything around it. The chart it reads from. The tools it is allowed to use. The double-checks before anything reaches a patient. Same clinician, different system around them, very different outcomes.

We already know this in pharmacy. The active ingredient is not the drug. Formulation, delivery, and the dispensing workflow determine whether the medicine actually works in the real world. Models are the active ingredient. Harnesses are the formulation.

The data is not subtle

A benchmark called Harness-Bench found that swapping only the harness, same AI model underneath, moved task performance by roughly 24 points. That is not a rounding error. That is the difference between a tool you trust and a tool you do not.

NVIDIA's paper pushes this further. Their framework posted 82.2% on SWE-bench Verified, a standard test of whether AI can complete real software engineering tasks. That is a few points above the best open alternatives. But the number that caught my attention was cost. They did it on about half the tokens, half the compute for better results, just from redesigning the scaffolding around the model.

Their six ideas, translated

1. Structured inputs and outputs. The AI fills out a validated form instead of writing free text. Like e-prescribing versus a handwritten script.
2. Working with live data instead of copies. The AI points at the record instead of pasting the whole chart into its memory.
3. Acting through code, not just clicks. It writes small programs to do multi-step work in one pass.
4. An editable playbook. The logic that runs the AI is plain code a team can inspect and change.
5. Durable memory. State that persists, instead of a conversation that forgets.
6. Self-managed context. The AI can see and clean up its own working memory, like a pharmacist clearing the counter between fills.

None of this is proven yet. Every number so far comes from NVIDIA's own tests, and the framework is a week old. But the direction is clear, and it is the same direction the whole field is moving.

What this means for healthcare

When a vendor pitches you an AI tool, the model name on the slide tells you maybe half the story. Ask about the harness. What checks the output? What can the system actually touch? What happens when it is wrong? Those answers live in the harness, and the harness is where safety lives too.

My rule of thumb, judge the system, not the model.

When you evaluate AI tools for your practice, what do you ask about beyond which model it runs?

Sources: NVIDIA Developer Blog, "Six Agent Harness Capabilities for Higher Model Performance." NOOA (NVIDIA-labs OO Agents), arXiv:2607.20709. Harness-Bench, arXiv:2605.27922.