Binder
In AI, the harness is the product
The model is interchangeable. The scaffold around it is what creates value.
Same model. One scaffold: 42% success on a benchmark. Different scaffold: 78%. The model didn’t change. Everything that matters — context loading, tool design, progressive disclosure, the architecture of the agent loop — lives in the harness. Build the harness.
The same principle applies to enterprise software
The question isn’t which LLM. The question is what workflows you’re automating, what domain knowledge you’re encoding, and what feedback loops you’re creating.
Domain expertise is the moat, not the model.
What this means in practice
- Don’t bet on the model. Every model gets cheaper, faster, and better. Anything you build that assumes a specific model’s quirks is a depreciating asset.
- Invest in the scaffold. Context, tools, evals, the loop. That’s where the compounding advantage lives.
- Domain over horizontal capability. The companies that win in vertical AI will be the ones who knew the domain before the model existed. Tacit knowledge is the durable layer.
The model is a commodity. What you build around it is not.