A software essay argues that agentic AI is shifting the main constraint in application development from writing code to judging whether the resulting system is correct. Its central claim is that deep knowledge of a business or technical domain has become more valuable as automated tools make implementation easier.
The author describes traditional programming as the transcription of a mental model. Before building payroll software, for example, a developer must understand deductions, garnishments and rate changes across pay periods. A transit application similarly depends on distinctions among routes, trips, schedules and real-world service. In this account, learning those rules was always the harder part, even when coding received more attention.
AI agents weaken the former connection between understanding and implementation because they can produce a working-looking program for someone who cannot write it independently. The essay contrasts two users. A logistics dispatcher, clinical coder or actuary may lack conventional software skills but can immediately recognize an impossible schedule or invalid claim. A capable generalist engineer can design and test a reliable service, yet may accept a plausible result that violates rules in an unfamiliar field.
The comparison does not establish that domain experts can deploy safe software without engineering support. Rather, it isolates two different layers of verification. Technical review asks whether code is secure, maintainable and operationally sound. Domain review asks whether outputs reflect the rules and realities they are supposed to model. An AI-generated billing rule can compile and pass the tests an engineer thought to write while still being wrong in practice.
The essay therefore identifies people with both capabilities as especially valuable. They can inspect generated implementation and also design meaningful tests grounded in real constraints. An hours-of-service test matters because someone knows the governing limit, not simply because the assertion is syntactically correct. The agent handles more of the translation from requirements into code; the human remains responsible for judgments at both layers.
One provocative part of the argument is that agents have shortened the domain expert's route into software more than the engineer's route into a domain. A dispatcher can ask a tool to produce code, while an engineer still needs time, observation and repeated exposure to acquire tacit operational knowledge. That asymmetry is presented as a change in career incentives, not a measured labor-market result.
The author's recommendation is for experienced engineers to build expertise in an industry, regulatory system, instrument or physical process. Whether coding skill has become “dramatically” less valuable remains an opinion. The more durable point is that faster code generation does not supply an independent oracle for correctness; organizations still need people able to define and recognize the right outcome.
