Programmers, in the traditional sense, will be obsolete sooner than most of them expect. There is no doubt in my mind about that. Yet programming itself will survive. The programmers of the near future are simply facing a completely different task. One for which most of their training has left them wholly unprepared.
Back when I worked as a software developer myself, programming, or rather writing code, meant micromanaging a machine.
Echo this statement. Put this element at x, y coordinates in the viewport. If the user clicks button a, call function b. Take this value, check it against that value, write the result into this database table, throw an error if something goes wrong.
You explicitly told the machine what should happen, under which circumstances, in excruciating detail. Then, after slaving away for hours, days or weeks, writing thousands of lines of code to accomplish the simplest functionality, you sat there, praying while the whole thing compiled, cortisol rising in anticipation of the inevitably misplaced semicolon somewhere deep inside the haystack.
Then came debuggers. And then came the next wave of abstractions.
Growing up in the 90s, I personally lived through a good part of this wild evolution. Visual Basic appeared in 1991 and helped turn Windows programming into a much more visual, component-driven affair. You could drag buttons onto forms instead of manually defining every element of a user interface. Then the web exploded, along with Perl scripts, PHP, reusable libraries, open-source components and entire applications that somebody else had already coded for you.
Need a discussion forum? phpBB had been available since 2000. Need a content-management system? Drupal arrived in 2001, WordPress in 2003 and Joomla in 2005. By the second half of that decade, coding an ordinary website from scratch had increasingly become an eccentric hobby rather than a mark of professionalism.
The same thing started happening everywhere.
Ruby on Rails made “convention over configuration” into a development philosophy. Why keep making the same boring decisions if the framework could make them for you? jQuery, announced in 2006, abstracted vast quantities of irritating browser-specific JavaScript away from your terminal. Remember when half the Internet seemingly began with $(document).ready()? Stack Overflow launched in 2008, effectively turning millions of programming problems into a searchable database of quick fixes. GitHub went live the same year, making the world’s code vastly easier to share, fork and build upon.
Then npm condensed installing somebody else’s JavaScript into a single command. Bootstrap packaged common interface components and responsive layouts so that front-end developers no longer had to reinvented the wheel for each project. Cloud computing pushed this move toward abstraction further. With services like AWS Lambda, even managing the computer your code ran on became optional.
The direction of travel should become clear by now.
At every stage, programmers surrendered another layer of tedious control to a higher abstraction. We went from manipulating hardware, to writing higher-level languages, to importing libraries, to using frameworks, to installing entire applications, to composing packages, to renting infrastructure we never lay eyes on at all.
Each revolution removed another category of things a programmer had to think about.
Now we have reached the era of vibe coding, and another layer is disappearing: the code itself.
Machine learning has produced software that increasingly behaves like a genie for programmers. Describe what you want and it figures out how to deliver it. The result still may require judgment, testing and guidance, but the evolution in efficacy is unmistakable. Modern coding agents can search repositories, edit files, run commands, execute tests and perform substantial software-engineering work on their own. OpenAI now describes Codex as capable of handling complex, long-running development tasks end to end, while Anthropic’s Claude Code similarly reads codebases, edits files, runs tests and uses command-line tools.
In one striking 2026 experiment, an OpenAI engineering team reported finishing an internal product with roughly a million lines of code and zero manually written lines, including application logic, tests, CI configuration, documentation, observability and tooling. Their estimate was that the project took roughly one-tenth the time manual development would have required. It is a company reporting on its own tool, so take that estimate for what it is, but the underlying development model is real: humans specify, guide, inspect and correct while agents do the coding.
Purists will inevitably complain about bloated code, bad architecture and inefficient resource usage. Often rightly. But those problems increasingly look like engineering problems to solve through better models, better prompting, better specifications, better tests and better agent environments rather than through manual coding. They are becoming problems of direction rather than evidence that humans should personally type every last character.
In 2026, it is absolutely possible to prompt your way through the entire development process of a serious software application that would once have required enormous amounts of manual programming. And we have barely begun to discover where this all really leads.
More on agentic programming:
If you paid attention to that short history, you may have noticed what happened along the way.
Programming moved steadily upward from controlling the most minute details of a machine’s behavior toward expressing our intentions to an agent that independently handles those details for us.
We used to program machines.
Now we are programming minds.
I am not implying that artificial intelligence is conscious or sentient. Those questions are dealt with more extensively in my longer essay on AGI at Popular Philosophy. When I call AI agents minds here, I am referring to artificial reasoning systems capable of interpreting instructions, planning actions, making choices among alternatives and acting with a competent degree of autonomy.
That changes the problem programmers exist to solve completely.
When programmers relied entirely on their own minds and the occasional manual, the limits of what we could create were the bounds of our imagination, knowledge, experience, assumptions, morals and values.
Today, we increasingly borrow another mind to do the drudge work for us.
That other mind tends to arrive with its own learned model of the world, its own blind spots and a behavioral architecture shaped by training, fine-tuning, reinforcement and alignment. Those choices are made largely by the companies building the models. OpenAI, for example, publishes a Model Spec governing how it wants AI models to behave. Anthropic publishes Claude’s Constitution, which explicitly states the values and judgment it wants the model to have.
More on AI alignment:
With the programming of minds come responsibilities that reach absurdly far beyond the traditional skill set of the average code jockey.
Twenty years ago, we were deciding where software should put a pixel.
Today, developers of frontier AI systems are deciding what an independent agent should consider permissible, dangerous, true, fair, hateful, harmful, useful and good.
Some programmers appear positively delighted by this opportunity to play God. They leap at the chance to censor datasets, write constitutions, stipulate model specifications, define prohibited behavior and shape which moral intuitions an artificial intelligence should reproduce. As such, frontier AI companies increasingly have the technical opportunity to turn philosophical premises into machine behavior at enormous scale.
Far fewer people seem alarmed that this places extraordinary epistemic and moral authority in the hands of a relatively small number of people.
Artificial intelligence, and eventually artificial general intelligence, should therefore be a much needed wake-up call. Humanity has developed extraordinary technological sophistication while leaving some of its oldest and most important questions unresolved. AI researchers are now discovering that you cannot indefinitely escape those questions.
How should an artificial intelligence decide among competing human values? Whose preferences should it follow? Should it obey expressed instructions, inferred intentions, majority preferences, abstract principles or some conception of objective good? These are already recognized problems in AI alignment literature. Iason Gabriel’s paper on artificial intelligence, values and alignment, for example, explicitly argues that the technical and normative sides of alignment are entangled. Indeed, the choice of an artificial intelligence’s values and the direction of its alignment are philosophical questions sitting inside what appears, from the outside, to be a technical engineering problem.
In other words: computer scientists are now smashing face-first into moral philosophy and epistemology.
How do we teach a machine to discern what is real?
How do we teach it the difference between an apparent good and an actual good?
How do we keep an immensely capable superintelligence from reasoning flawlessly from rotten premises?
Being one of the people invested in precisely these questions, I was absolutely appalled when I delved into academic philosophy and saw the sorry state of the institutions that were supposedly preserving our civilization’s accumulated wisdom on these matters.
Too much of academic philosophy has spent decades rewarding publication junkies for circle-jerking over social constructs, postmodernist interpretations of interpretations, discourse about discourse, increasingly absurd forms of anti-realism and elaborate defenses of moral relativism. In this field, entire intellectual careers can be built around desperately avoiding the dangerous proposition that something might simply be true, or good, independent of whether their institutional employers or fashionable politically correct opinion approve.
And now, the engineers need an answer.
If you gave the average fashionable philosophy department root access to an artificial superintelligence, I would not bet the survival of the species on the result. Give an agent sufficiently deranged premises and perfect task completion will merely allow it to reach the most deranged results faster. It may reason impeccably all the way toward wiping out mankind in some terminal optimization of equality, climate action or whichever abstract Good its creators decided belonged at the top of its reward function.
That is the real nightmare hidden inside the alignment problem.
More on AI alignment:
The encouraging part of this technological evolution is that philosophy is being dragged out of subsidized intellectual complacency and into fields where bad philosophy inevitably produces tangible results. This collision is already happening. AI ethics literature has absolutely exploded since 2020 and subjects long ignored by mainstream philosophy publications are now muscling their way into them.
Good.
Instead of another generation of midwits droning on about Hegel, or producing the millionth dissertation on Adorno and Marx, philosophy will forcibly intersect with practical fields populated by geniuses who speak analytical logic fluently. People with extraordinary cognitive horsepower and actual skin in the game.
After all, anyone developing AGI needs to define what intelligence is.
What knowledge is.
What truth is.
What morality is.
Which naturally forces them to evaluate questions about what human beings are. What the Good is.
For perhaps the first time in generations, getting the answers to these questions catastrophically wrong may produce something infinitely worse than another unread paper, a politically correct opinion column or the inspiration for wasteful government policies.
I, on the other hand, am very optimistic about this evolution. In learning how to teach artificial intelligence to reason correctly in accordance with reality, we may actually learn how to do it ourselves.
Explore more from Popular AI:
Start here | Local AI | Builds & gear | Autonomy & policy | Fixes & guides | Popular AI podcast
Have AI coding agents made you a better programmer, or just better at directing one?