Is vibe coding the end of programming?

The claim has a 70-year history. 2026 is the first time it has teeth.

"The end of programming" is a recurring genre. FORTRAN's automatic-coding era in the 1950s promised to move software away from line-by-line coding; model-driven engineering, program synthesis, and low-code platforms each revived the promise. Matt Welsh's January 2023 Communications of the ACM essay "The End of Programming" made it explicit: "the end of classical computer science is coming, and most of us are dinosaurs waiting for the meteor to hit" — software would be trained, not written, and "programming will be obsolete." For two years it read as provocation. In 2026 it reads as a position statement in a real argument.

What changed: agents started shipping systems, not snippets. Paul Dix's August 2026 essay "The end of programming" put the fresh evidence at the center — Bun 1.4's rewrite from Zig into Rust, more than a million new lines produced by agents working under one developer, Jarred Sumner, with a pre-release frontier model and effectively unlimited tokens. Eleven days to the working rewrite, then months of agent-driven refinement before release — software running on millions of developer machines, most of which no human ever read line by line. Dix's conclusion is the strong form of the claim: "we will have more running, production software written by AIs than written by humans. It'll be the end of programming."

And the person who named vibe coding moved on. Andrej Karpathy coined the term in February 2025, reported ~80% of his own code AI-generated by December 2025, and by April 2026 (at Sequoia's AI Ascent) declared vibe coding passé in favor of agentic engineering — "the professional discipline of coordinating fallible agents while preserving correctness, security, taste, and maintainability." His own framing of the relationship: "vibe coding raises the floor. Agentic engineering is about extrapolating the ceiling."

The evidence for "yes" — taken seriously

• The artifact is no longer human-scale. Bun 1.4's ~1M-line agent-written Rust rewrite (Aug 2026) is production, not demo — the strongest existence proof that "humans write the code" is already false at the frontier.
• The labs already work this way. Anthropic and OpenAI engineers publicly ship dozens-to-hundreds of PRs a week and describe building "loops" — systems, prompts, and verification tooling — rather than reviewing every line. Dix's read: they are review-at-the-result, not review-at-the-code.
• The volume shift is measurable. GitHub-reported stats put roughly 46% of new code on the platform AI-generated; a quarter of YC's 2025 batch ran codebases ~95% AI-generated.
• Cost is collapsing. Frontier-model prices fell hard through 2025-2026; the marginal cost of producing working software-shaped output keeps dropping toward zero.

The evidence against — also taken seriously

• The only randomized controlled trial says experts got slower. METR's RCT (16 experienced open-source developers, 246 real tasks, early-2025 tools): tasks took 19% longer with AI — while the same developers believed they were 20% faster. METR flags the result as historical to early-2025 tooling, but no published RCT has yet shown a measured speedup for experts on mature codebases.
• The quality tax replicates. The September 2026 three-way study (arXiv 2609.09560) measured vibe coding 27% faster than traditional coding — with lower maintainability indices and more security vulnerabilities. The ACM/IEEE ICSE-SEIP grey-literature review (Jul 2026) found practitioners describing output as "fast but flawed," with QA the first casualty.
• Trust is falling while use rises. Stack Overflow's 49,000-developer 2025 survey: 84% using or planning to use AI, but trust in accuracy fell from 40% to 29% in one year, and 66% report frustration with "almost right" AI code.
• Comprehension cannot be delegated. The pedagogy study behind the "epistemic debt" paper (arXiv 2602.20206): unrestricted-AI novices suffered a 77% failure rate on a later AI-blackout maintenance task vs 39% with scaffolded AI use. If nobody can read the code, nobody can maintain it when the model can't.
• The correctness-oracle problem is structural. Matt Yancer's February 2026 rebuttal is the sharpest short version: generation is cheap, correctness is the job; "does this program do what I meant?" has no universal button — tests, contracts, types, staged rollout, and rollback are still engineering, whether a human or an agent typed the code.

The Princeton framing (Narayanan & Kapoor, 2026) reconciles the two columns with the decide-execute-deliver sandwich: AI compressed the middle of software work (the execution), but the two ends — deciding what to build and being accountable for what ships — are unchanged. CMU's Shaw, Hilton & Fairbanks in IEEE Software (Mar 2026) land the same place: success now depends on "reading, critiquing, and modifying code rather than writing programs."

The jobs data: the word "programmer" is doing all the work

The end-of-programming debate keeps citing the same employment numbers selectively. Put them next to each other and the actual story appears — a title split, not an extinction:

Read honestly: the bottom rung broke; the profession grew. AI absorbed the well-specified junior tasks that were also how juniors became seniors — the AIES 2026 study ("Who Will Become the Next Senior?") documents the apprenticeship ladder eroding in interviews across Korea and the US, where seniors-plus-AI now absorb the debugging and documentation work juniors used to learn on. That is the real cost the "end of programming" headlines skip: not mass replacement, but a pipeline problem. Meanwhile total developer employment grew, senior demand rose, and the highest-exposure companies kept hiring.

The honest forecast is neither "coding is dead" nor "nothing changes": routine implementation is commoditizing fast, judgment (specification, review, verification, architecture) is appreciating, and the field renamed its own future — from vibe coding to agentic engineering. "The end of programming" is true as a description of what the work is and false as a prediction about who is employed to do it.

What ends, what survives

So is vibe coding the end of programming?

The precise 2026 answer: it is the end of programming as a typing job, and the beginning of programming as a direction-and-verification discipline. The people shipping the most software in 2026 still "program" — they specify intent, design the system's boundaries, build the evals and harnesses that catch regressions at machine speed, and own the decision to deploy. Karpathy's four practices (detailed specs, critical diff review, automated eval loops, agent-permission management) are the job now; eval design is the one he calls most commonly underdeveloped. The history of the field says every abstraction wave — assemblers, compilers, frameworks, low-code — was mistaken for the end of programming, and each time demand for software expanded instead. The difference this time is distribution: for the first time the bottom of the ladder is compressed at the same moment the ceiling rises. That combination is new. "The end of programming" is the wrong name for it; "the great reshuffle" is closer.

If you want the mechanics of the new job rather than the obituary of the old one, work through the vibe coding guide, the controlled comparison in vibe coding vs traditional coding, the production discipline in agentic engineering, and the numbers in vibe coding statistics.

How VibeFuse fits the post-typing workflow

VibeFuse is the first ever free widget-based AI harness: Claude Code, OpenAI Codex CLI, Gemini CLI, Cursor Agent, and Qwen run as live widgets on one Windows canvas, with terminals, files, and a browser beside them. If the future is directing fallible agents, the harness is where the direction and verification actually happen — you watch every agent action live, inspect diffs while the context is still on screen, and keep named session records as the audit trail that "who approved this change" questions require. It runs local/offline, so your code and build activity never pass through a third party's cloud, and it works in any app.

The verification layer is open too: the open-source marketplace lets builders make money selling widgets, skills, and styling packs at an 80% revenue share — review checklists, secrets scanners, and test-coverage widgets that other builders ship into the same loop. The end of typing does not mean the end of judgment; it means judgment gets tooling, and the tooling is something you can own and sell rather than rent.

Is vibe coding the end of programming — FAQ