Software is Dead: The Harness That Turns AI into an Expert

The concept that software is dead and programming, as understood for three decades, is becoming obsolete has taken hold among those building digital products. This shift is underpinned by the directions being taken by YCombinator and the views of veteran Robert C. Martin, known as Uncle Bob. The formula summarizing this argument is attributed to LangChain: agent equals model plus harness.

What is an AI harness and why is it presented as the end of software

The harness is a layer of software and instructions that surrounds a language model, forcing it to behave predictably. The sequence is well-known: prompt engineering dominated between 2022 and 2024, when simply writing better instructions sufficed; context engineering took over in 2025; and the harness arrived at the beginning of 2026. The starting diagnosis is uncomfortable, because a prompt is a suggestion the model can ignore, and an entire industry rests upon that fragility.

Its commercial application is explained in two sentences. A generalist model—cited as an example, ChatGPT Astra—is taken and layered with files and rules until it becomes a specialist: one in legal domain, another in medicine, another in sports history. What is sold is not the model, but the wrapper.

The Achilles' Heel: A Statistical System That Hallucinates

The engine of this entire process is not deterministic. It returns statistical truths, not rational ones, which necessitates building layers of control. With precise instructions and revisions, it works; without them, the result is a mass of code that must be reworked. The comparison that keeps resurfacing is that of a slot machine: it hits often and fails without warning.

There are two bottlenecks that product literature tends to overlook. The first is the context window: if a project is more than just a form, instructions accumulate and the model loses accuracy. The second is Retrieval-Augmented Generation (RAG), sold as a magic solution when it depends on the internal search retrieving the correct fragment; if it doesn't, the model won't see it. Hence, much of the sector maintains that the tool only performs with close supervision and highly specialized agents.

Retiring programmers and discovering that auditing the code costs more

The uncomfortable economic argument arises when looking at the bill. Several companies that outsourced development to AI models are now facing trouble: the generated code contains errors, it must be reviewed line by line, and the volume produced multiplies what was written before. No one does this review easily, and the cost of auditing exceeds the saved labor. Furthermore, it is maintained, without official confirmation, that some Linux distributions have closed their repositories to AI-generated code.

Here lies the fundamental contradiction. In a software company, it is not just the capacity of the coder that counts, but their knowledge of the architecture, procedures, and system history. That asset cannot be replaced with a harness.

The 'Sovereign AI Stack': Buying GPUs to be Independent

It's not all smoke. Latham & Watkins, the second-largest US law firm by revenue, has purchased servers with Nvidia GPUs to host AI systems on its own infrastructure, according to the Financial Times. The coined concept is sovereign AI stack: open-weights model, proprietary dataset, and dedicated infrastructure.

It is important to choose words carefully. Fine-tuning a model with proprietary data is expensive; it requires specialized personnel and many human evaluators. And not everything called training is training. The most common technique, low-rank adjustment, improves the behavior of an existing model but does not reinvent it: you train the dog, and it remains a dog.

The Hurdle AI Can't Clear: Selling

Multiplying technical production capacity by twenty collides with a wall that is not technological. Those who no longer need programmers discover they still need clients. And selling is the expensive part: without sales, the thousand previous passes are worthless. Marketing is recycled into a new terrain—getting the AI itself to recommend a product. People are already asking assistants what to buy, and this opens up a discipline—optimization for generative models—that involves persuading the intermediary rather than the consumer.

Spinners, Weavers, and Typists: The Troubling Precedent

The historical parallel repeats itself: mechanization wiped out spinners and weavers, and no one missed them as a labor category. Applied to the present day, the job of typing will be reduced to programming critical and highly specific systems. Anyone who thinks they can spend their entire career writing application code is signing their own termination notice.

And this is where the analysis stalls. If AI smoothly replaces what persuades and struggles with what must function; if it makes writing code cheaper but verifying it more expensive; if no one has compared the auditing cost with the salary of the programmer being replaced, then the obituary for software arrives, at minimum, prematurely. The full bill is missing.