What is AI programming called: a dozen names, two chosen by the market

The entire history of programming is a ladder of abstractions: every rung hid a detail previously considered "the essence of the profession." Assembly hid numeric operation codes, the compiler hid registers, Python hid memory, and the agent hides syntax. Natural language is not the end of programming but its next floor: Karpathy framed this as "Software 3.0" - the prompt becomes the program, and the LLM its interpreter.

The same thing happened at every rung: first, "this is not real programming," then, "well, it will do for prototypes," then it becomes the default - and the marker word dies of irrelevance. Nobody says "compiler programming": it is simply programming.

Hence the prediction: terms of a transitional era dissolve when the phenomenon becomes the default. "Vibecoding" and "agentic engineering" are names for the bridge between eras; once the bridge is crossed, only "programming" will remain - at the most powerful available level of abstraction. The map of terms below is a snapshot of the moment while the phenomenon still has names.

History came full circle: Karpathy coined vibe coding as a "shower tweet" about weekend projects - and a year later introduced agentic engineering himself, because his own word was not enough for serious work. The author of the term was the first to outgrow it.

Russian-language segment

A telling episode in the Russian-language segment: the Habr longread "How vibecoding drove me to depression, then led me to vibe engineering" repeated Willison's trajectory - Russian-language arguments follow the same path as English ones, lagging by a couple of months.

For every new AI coding term, we establish a Russian name: the verdict, date, and measurement it rests on (search suggestions, frequencies, practitioners' language). Argue over what is correct - send a link to this table. We do not discard alternative names: a term card can be found by any spelling.

The glossary grows alongside the term: it appears in the feed - a row in the table; it takes hold in practitioners' speech - its own page; search demand appears - a full concept card, like loops.

The mistake in most arguments is treating vibecoding and agentic engineering as competitors. They are not two camps but two ends of one scale, differentiated not by the tool (everyone uses agents), but by the amount of engineering around the result:

the left lifts the floor: anyone can build software · the right lifts the ceiling: quality at agentic speed · axis - Karpathy's framing (Sequoia Ascent 2026); Willison's criterion: do you read the diffs

The mockery has an honest side: carelessness is a real problem, and the critics' numbers are real. METR uncovered the most uncomfortable fact in a rigorous experiment: experienced developers using AI tools took 19% longer - while believing they had sped up by 20%. But the mockery itself is not an argument but a phase: this is how every new floor of abstraction has been received.

This pattern was described long before AI: Everett Rogers' innovation adoption curve (1962) divides people into waves - from innovators to laggards, and laggards - the last 16% - can be recognized by one trait: they laugh at what will become their own work tool a few years later. "Mom's vibecoder" is "Real Programmers Don’t Use Pascal," republished for Zoomers: the 1983 manifesto now reads like a historical anecdote, and the abstraction ladder's score is 5:0. The dividing line is not between "writes by hand / generates with an agent" (everyone uses agents in 2026, including those who mock them), but between "takes responsibility for the result / does not": reviews, tests, and responsibility distinguish an engineer from a careless beginner on any floor - whether in assembly or English.

The analysis must withstand its strongest opponent - and that is not an anonymous commenter, but Edsger Dijkstra. Back in the late 1970s, he wrote an essay with the telling title - “On the foolishness of ‘programming in natural language’” (EWD667):

Dijkstra was right - for his century. As long as no one but humans could keep the imprecision of natural language in check, formality remained the only defense against nonsense. LLMs changed exactly that: for the first time, a machine can work with an imprecise formulation - it will ask follow-up questions, fill in context, offer an interpretation. But note that not has been automated: the responsibility for ensuring that what gets built is what is needed and works as it should. What remains of Dijkstra's case has not vanished - it gained a name: agentic engineering.