AI is giving organizations more power than ever before. Anyone can now produce software, agents, and automations in hours. Far fewer organizations can say what all of it is for, who owns it, or how it will be kept under control. In my country (I am Dutch), there is a saying: het kind van de rekening. It means that someone has to carry the cost as a burden for someone else. That unfortunate person in the age of AI-generated software will be whoever has to operate the result.
It’s no surprise that coding agents have collapsed the cost of building. But as Charlotte Fleming of Octopus Deploy put it in her platform engineering research, “Building has become easier, but maintaining what gets built hasn’t.” Her survey of 379 platform professionals found that teams building mostly from scratch were about 3.5 times more likely to report that their platform lacked a clear strategy. This shows that building by default, without a well-defined reason, goes hand in hand with having no strategy.
Researchers (“Debt Behind the AI Boom: A Large-Scale Empirical Study of AI-Generated Code in the Wild”) examined software projects to measure what happens after AI-written code goes into production. In all five studied AI coding tools, at least 15% of the code changes introduced new problems that someone had to fix later. The worrying part? AI created more correctness and security issues than it solved. This led to the following conclusion: AI-generated code brings long-term maintenance costs into projects, so it needs stronger quality control. GitClear’s analysis of 623 million code changes adds that code duplication has risen 81% compared with pre-AI levels, while code reuse has fallen. Every copy is one more place where a fix must be made (“The Maintainability Gap: AI Code Quality in 2026”).
While code can be generated in seconds, the context around it cannot be. Why does a system exist? What does its data mean? Which limits must it respect, and who is responsible for it? That knowledge rarely gets written down, so no model can generate it. And without it, every new agent, integration, and automation adds complexity that nobody fully understands.
Most likely this debt won’t appear as a failed build, but as an outage nobody can explain or as an AI agent acting confidently on data it has misread. The worst-case scenario is a system nobody dares to change, because nobody knows what depends on it. In factories, utilities, and buildings where systems run for 20 years and mistakes have physical consequences, this debt will grow the fastest and will cost the most.
But not all is lost—help is on the way. For 30 years, the question was how to build, but that’s now largely answered. The difficult question has become how to operate what AI builds and does, safely and over time. That is why harnesses are on the rise. (At the time of this writing, harnesses are not known by the larger audience.)
What is a harness? It’s everything around the model: the context it receives, the tools it may use, the boundaries it must respect, the checks that verify its output, and the records that show what it did and why. The model provides intelligence. The harness turns that intelligence into something an organization can trust, govern, and maintain.
Let’s look at the progression. Prompt engineering is about phrasing one good instruction. Agent orchestration is about coordinating many agents across a workflow. But orchestrating agents that lack shared meaning and enforced boundaries only multiplies the debt. The next wave or discipline is harness engineering. This is deliberately designing, building, and maintaining the environment agents work in, as a lasting asset rather than an afterthought. Good examples of harnesses are the frameworks the AI companies like providing, such as Claude Code for coding or Claude Cowork for office tasks.
There won’t be one harness for everything. While models generalize, harnesses will specialize, because the meaning they carry belongs to a domain. A water utility’s harness must know its pumping stations, its discharge permits, and what a pressure drop at 3 a.m. means. A factory’s harness must know its machines, its shift patterns, and which safety interlocks no agent may ever override. So, we will see many harnesses, built for different industries and purposes by the specialists who know those domains best. The models may become a commodity. The harnesses will not.
Prompts are disposable. Orchestrations change with each project. (I heard someone saying that after 3 months, only 2% of AI agents still fully function.) The harness is what remains, and it is where an organization records what the mission is, what each agent is for, and where its authority ends. A harness turns raw strength into useful work, pulling in the right direction. That is why the winners of the next decade will be those who engineer the best harness, so they are not left paying the bill.