AI coding agents have changed software development faster than any earlier shift, and one veteran developer says the industry is underestimating what it is giving up. The argument borrows from Frank Herbert’s Dune, where a rare substance called spice extends life and expands consciousness, but also creates dependence and concentrates power.
The author says they use AI coding agents for much of their working day and finds them almost intoxicating. Tasks that once took days now take hours, and projects that sat on an ideas list can become working software in an afternoon.
The blind spot of experienced developers
Some of the most accomplished traditional programmers are now leading the push toward fully agentic work. They say English is the new programming language and that syntax is dead.
The author sees a blind spot. An experienced developer brings decades of knowledge to every agent session: how computers work, how databases, networks and concurrency behave, which questions to ask and which answers to distrust. That knowledge is so ingrained they may stop noticing they use it. Building an app without writing a line of code can feel like proof that programming knowledge is unnecessary, when it guided nearly every decision.
There is a gap, the author says, between no longer needing to exercise a skill and never having acquired it. In Dune, humans trained as Mentats after thinking machines were banned. Experienced engineers resemble modern Mentats, with AI amplifying abilities they already have.
English isn’t a programming language
A common defense says models are the latest layer of abstraction, from machine code to assembly to C to Python and now English. The author disagrees. Python removes complexity but still demands precision: functions have inputs and outputs, conditionals follow rules, and loops force you to think about state and termination.
Natural language has no such guarantees. When you prompt a model, it does what it statistically guesses you meant. For most of its history, the author notes, programming was about removing ambiguity. Now it is shifting from specifying behavior to eliciting it.
The lost apprenticeship
The industry has long worked as an informal apprenticeship. Juniors handled simple tasks under supervision and gradually took on harder work until they were the reviewers. Much of that entry-level work now goes to agents, and junior hiring is struggling.
Hiring only experienced developers who can do a team’s work with tokens makes sense for one company, the author argues, but could be damaging for the industry. They compare it to training airline captains by having them supervise autopilot from day one. They hope AI will offer better ways to build expertise, but say the old system is being dismantled before a replacement exists.
Is the architect safe?
Many senior engineers say agents will handle implementation while humans keep architecture and design. The author finds that convenient, since it is how many seniors prefer to spend their time. Architecture involves incomplete information and office politics, but those explain current AI limits and don’t prove lasting human superiority. If models keep improving, architecture may not stay human. If they don’t, abandoning implementation skills looks premature.
Who controls the spice?
The author is also puzzled that longtime open source advocates, who taught them to distrust black boxes and vendor lock-in, now embrace proprietary AI. The economics may not hold: models might stop getting cheaper, or providers might stop selling access at reasonable prices. Open-weight models and local inference help, but training frontier models is out of reach for most individuals and institutions.
In Dune, power comes from controlling the supply. The author’s phrasing is that developers are “exchanging knowledge we possess for intelligence we rent.”
Ghosts in the model
The author also compares agents to Duncan Idaho, a character repeatedly revived as a ghola. Inside the models are traces of generations of programmers: the people who built Unix, designed languages, maintained obscure libraries and answered strangers’ questions online.
What the industry should decide next
The piece is an opinion, not a forecast, but it poses concrete questions. Who trains the next generation of senior engineers? What happens if frontier pricing changes? How much hands-on skill should teams keep? Companies that lean on AI coding agents will need answers before the junior pipeline runs dry.
Source: Will Vincent