Here's Where I Am At

Jul 30, 2026

Not sure about you, but I’ve been struggling a lot to figure out how things are shifting in the software industry. I think I’ve made some progress, and here’s where I am currently at:

• Software companies are shifting to selling work, not software. Customers never really wanted the software; they just wanted the results that could be produced by that software. Those results are produced by the work done by a person using the software. That work could never be offered at scale. Agentic systems make offering work directly practical. Also, as the price of building software goes down, competition rises. So selling the work is a natural path to be closer to customers and survive competition.
• What is the thing that allows selling work at scale? Claude Code is software that sells programming work at scale. Claude Code is an LLM model wrapped by an engineering layer that gives the model agency to execute the programming work. That engineering layer is called harness. The thing is the harness on top of the model.
• No matter how good models become. A harness will always be needed. Because the harness is the interface with the external world. It is the thing that gives the model context of its environment, allowing it to make decisions and to know when it is right or wrong and how to adjust.
• There’s no such thing as AGI. Maybe there’s a world where a general model exists, but a general model without a harness is useless. Having an AGI would mean a general harness that would give a model agency over all kinds of work is possible. But the set of all kinds of work is infinite.
• If companies are shifting to selling work, and the harness is the thing that allows them to do that, companies are shifting to the business of building and selling harnesses.
• The product is the harness. The challenge is figuring out what that harness looks like for a particular work being automated. I don’t think an AI will automatically give an answer here. Figuring this out requires expertise in the domain being automated and good judgment. Also, the human on the other side that is buying your harness has a particular way of seeing the world and how that work should be automated. So, even within a domain, many harnesses with different tastes, delivering different results, will emerge. And there’s no end to this. Claude Code is not a finished harness. It will keep trying to get to the bottom of what delivering programming work means. But there’s no bottom.
• Building harnesses is building software. Because we had none of these kinds of software before, it means the software industry will grow, and more software engineers will be needed. Hence, the rise of Harness Engineering and the AI Engineer.
• So the software developer who works on building harnesses will bring the good old software practices plus whatever new practices emerge from harness engineering.
• You can say we already have pretty good harnesses that sell programming work. Software development is a layer on top of programming, and software companies are trying to figure out (at the same time) the harness that allows them to develop software at scale while figuring out the harness to sell to customers. 
• The amount of Human-In-The-Loop necessary is inversely proportional to how good the harness is. But there will always be the need for HITL, even if tiny compared to the past.
• The competition to be the first in the segment will be ruthless. The ones with proprietary data have an edge. The harness can be better adjusted to that data.
• Nontechnical people will try to build their own software to make themselves more productive. But if there’s a harness out there that does that work already, it might be that the company those people work for might prefer buying the harness instead

The software industry has all different kinds of software. The kind of software I talk about here is the one that people hire to produce work. I apologize for the amount of the word harness I use (I am tired myself), but this is where I currently am. If you have thoughts on how this is going and where I might be mistaken, I’d love to hear them.