AI is the best teacher. Most people can't use it.

The teacher who never sleeps

Andrej Karpathy named "vibe coding" in February 2025. By 2026 he had retired the term, replacing it with "agentic engineering." Three months earlier, Anthropic — whose Claude Code now writes an estimated 4% of all public GitHub commits — quietly published a study showing AI assistance made programmers 17% worse at understanding the libraries they had just used. The hype hasn't broken. The classroom has.

This is an essay, not a benchmark. The thesis is short: AI is the most accessible teacher we have ever built, and most people get worse using it. Not because the model is wrong. Because the user can't amplify thought through it. Vibe coding was the gateway. Context engineering is the literacy. And our education has not caught up.

Karpathy retracts: from vibe coding to agentic engineering

Karpathy's original tweet was a mood: "fully give in to the vibes, embrace exponentials, and forget that the code even exists." Twelve months later, the same author replaced it with this: "review with the same rigor you'd apply to a human teammate's PR. If you can't explain what a module does, it doesn't go in."

That is not a small edit. It is a public concession that the original framing dropped a discipline the work always needed. Anthropic's CEO Dario Amodei told CNN in early 2025 that AI would write all the code in three to six months. Six months after that claim, Anthropic's own researchers published a paper showing devs who relied on the assistant scored 17% worse on conceptual questions about libraries they had just used.

These are not opposing facts. They are the same fact stated twice. AI generates code faster than any human team. That speed exposes — it does not create — the gap between people who can verify what they shipped and people who cannot.

The 80 percent wall

A DEV Community audit of 1,764 vibe-coded apps in 2025 found 453 — about 26% — with critical vulnerabilities. Roughly 80% repeated the same Supabase row-level-security mistake:

-- The bug that ships in 4 out of 5 vibe-coded apps.
CREATE POLICY "user_read" ON profiles
  FOR SELECT
  USING (auth.role() = 'authenticated');
-- "authenticated" means "any logged-in user." Not "the owner."
-- Every account holder can now read every other account holder's row.

A non-developer running Lovable or Bolt does not know this bug exists. The AI does not flag it. The deploy succeeds. The app ships. The breach is silent until someone notices.

Now hold that against two real cases. Plinq, built by a Brazilian non-developer named Sabrine Matos, shipped a women's-safety app in 45 days on Lovable plus Supabase and reached $456K ARR in three months. Friend.com, built by a well-funded team with $8M raised and a $1.8M domain, shipped a fragile chatbot and died publicly.

The difference between them is not coding skill. Both produced AI-generated code. One verified what the code did against a sharp problem and a sharp user; the other bought distribution before defining either. Speed without verification is survivorship, not method. The 80% wall is real. It just isn't visible from inside the build.

The thinking-amplification gap

The same tools that make seniors faster make average users slower. One 2025 productivity study found the average AI-coding-tool user posted a 19% productivity decrease. Anthropic's research found the 17% drop in conceptual understanding. A 2025 LeadDev survey found 54% of engineering leaders are planning to hire fewer juniors because "AI handles the easy work."

That last number is the dangerous one. The pipeline that traditionally cleaned up bad code — juniors learning to read, debug, and refactor under senior review — is being closed at exactly the moment when AI is producing more bad code than any junior army could clean. Forrester now expects 75% of tech leaders to face moderate-to-severe technical debt by 2026. A separate study of 8.1 million pull requests found tech debt rose 30 to 41% in repos that adopted AI coding tools.

What separates the senior with a 10× claim from the average user with a 19% loss is not access to tools. Both have Cursor. Both have Claude Code. The variable is the user's capacity to amplify their own thinking through the model: to brief it with intent, to spot when it is wrong, to redirect it before the wrong path becomes a buried bug. Most people cannot do this on day one. Some never learn.

This is the gap our education does not name. We teach syntax to people who will never write a for loop unassisted. We teach prompts to people who will be writing specifications. We teach tools that will be obsolete in six months to a market that needs literacy for the next decade.

Context engineering is the new literacy

The industry has already named the replacement. In 2026 surveys, 82% of IT and data leaders say prompt engineering alone is insufficient to power AI at scale. 95% of data teams plan to invest in context-engineering training in 2026. A European fintech that rolled it out across 50 developers saw a 40% drop in code-review corrections, 35% better security compliance, and 28% less time spent reformulating queries to its AI agents.

Context engineering is not a hot tool. It is a discipline:

• Prompt engineering asks: how do I phrase one request?
• Context engineering asks: what does the agent need to know, remember, and reach, every time it acts on this codebase?

A working context for a coding agent has four layers:
  1. Project rules     CLAUDE.md / AGENTS.md / .cursor/rules
  2. Design intent     a spec doc the agent reads before writing
  3. Tools             MCP servers, scripts, type signatures it calls
  4. Feedback          tests, lint, gap analysis, runtime logs

One framing has stuck: "if you're building agents in 2026, you are a data engineer — an architect of context." That sentence does not survive in a curriculum that still teaches "how to write a good prompt." It needs a curriculum that teaches how to design the environment the agent operates inside.

bkit: one attempt to operationalize the discipline

I built bkit because the gap was personal. I wanted my partner — who has never written a git rebase in her life — to ship a working web app and have it survive the second feature. The first try collapsed at the 80% wall. The second try did too. The third try, with a spec doc, a verification loop, and a fixed cadence, shipped.

bkit is one operationalization of the principles above:

• 43 skills + 36 agents + 142 lib modules + 21-event hook system. Not a chatbot wrapper. A harness.
• A gap-detector agent measures Match Rate between the design document and the shipped code. If Match Rate < 90%, an iterator runs up to 5 cycles to close the gap before the user is even asked for review.
• Three philosophies that name the gap as methodology, not skill. Automation First: PDCA is applied even when the user has never heard of PDCA. No Guessing: if the spec doesn't say it, the agent asks; it does not hallucinate. Docs = Code: the design and the implementation are kept in sync by the harness, not by hope.

I am not pitching this. I am reporting it. bkit is one attempt — among many that need to exist — at turning the discipline juniors used to absorb across years of mentorship into something a harness enforces in minutes. The principles do not require bkit. They require a curriculum that teaches them as the default mode of working with an AI agent.

Korea's consumption paradox

Anthropic's leadership pointed at Korea twice between 2025 and 2026. Korea is top-5 globally in Claude usage and top-5 per capita. The single highest individual Claude Code user worldwide is a Korean software engineer. Anthropic opened a Gangnam office. Korean Claude Code monthly active users grew roughly 6× in four months; Korean revenue grew 10× in a year.

And yet every dominant tool — Cursor, Claude Code, Copilot, Lovable, v0 — is foreign. 패스트캠퍼스 sells a 17-hour "바이브코딩 바이블" package. 노마드코더 markets directly to "바이브 코딩 실패자" — vibe-coding failures. A 더스쿠프 survey reported 53.1% of 571 Korean office workers have done a side project. The demand is overwhelming.

What is being taught is still mostly which prompt to type and which button to click. The half-life of an "AI coding" course on Inflearn has collapsed to six to twelve months. The only sustained domestic methodology contribution visible right now is Toss's "harness engineering" framing — a Korean fintech publishing what it learned about raising the floor of org productivity through a structured AI harness, not a faster model.

That is the export opportunity hidden in the consumption paradox. Korea has both the user base and the operator depth. What it does not yet have is education that treats context engineering as literacy before tool-of-the-month — that prepares Korean builders to ship methodology, not just consume infrastructure.

A curriculum for the next decade

If AI is the best teacher, what does the curriculum look like? Not as a course outline — as a literacy stack:

• Day 1: Specification before prompt. Three lines on a sticky note: what the system must do, what it must never do, what proves it's done. The non-developer who learns this on day one ships safer code than the senior who skips it. This is the only skill that compounds with every model upgrade.
• Week 1: Design-doc fluency. Read and write one-page design documents that machines can re-read. Not for the architect — for the agent. A spec the agent re-loads is a spec that survives a hundred AI rewrites.
• Month 1: Feedback-loop literacy. Match Rate, gap analysis, test signal, runtime logs. Treat the AI's output the way a senior treats a junior's PR: respect and skepticism in equal measure. Build the muscle to say "this is wrong" before the bug ships.
• Quarter 1: Harness composition. Skills, agents, state, hooks. Not as bkit features — as a generic literacy. Pick a tool, but learn the pattern. The tool will be replaced; the pattern will not.

The endpoint of this curriculum is not a generation of vibe coders who ship faster. It is a generation that uses AI the way the best students have always used the best teachers: as a force multiplier on thought they were already doing the work of having. Anthropic dropped a hint when it eased "vibe coding" out of its own product framing in early 2026. Karpathy dropped a louder one when he replaced the term he had coined. The signal is the same: speed is not the bottleneck anymore. The literacy is.

Summary

• Karpathy retired "vibe coding" within a year of coining it; Anthropic's own research found AI assistance worsens conceptual understanding by 17%. The framing was always provisional.
• 26% of audited vibe-coded apps ship critical vulnerabilities, and 80% repeat the same RLS bug. Speed without verification is survivorship, not method.
• The gap between a senior with a 10× claim and an average user with a 19% productivity loss is not access to tools. It is the capacity to amplify thought through the model.
• 82% of IT leaders now say prompt engineering alone is insufficient. Context engineering — the design of the agent's persistent environment — is the literacy that closes the gap.
• bkit is one operationalization (spec-first, verify-against-spec, automate the loop). The principles do not require bkit; they require a curriculum that teaches them as default.
• Korea over-consumes AI coding tools but under-produces methodology. The export opportunity is teaching context engineering as literacy before the next foreign tool ships.

For adjacent reading: see vibe coding with context engineering for the 29-day execution proof, intent is the new craft for the essay on what craft means now, and bkit + harness engineering for the technical positioning behind the case study above.

FAQ

Is vibe coding dead?

No, but Karpathy himself moved past the term. The framing was always provisional; "agentic engineering" names the discipline the original wording dismissed.

Do non-developers need to learn programming fundamentals first?

They need to learn specification and verification, not syntax. The skills that compound with AI are the ones AI cannot yet do for you: deciding what "correct" means.

What is context engineering, in one sentence?

Designing the persistent information environment — docs, rules, tools, memory — that an AI agent operates inside, so its outputs become reproducible instead of accidental.

Is bkit required to do any of this?

No. bkit is one open-source operationalization. The principles (spec-first, verify-against-spec, automate the loop) work in plain CLAUDE.md with a checklist; bkit just enforces them by default.

Why does Korea matter to this story?

Korea is the world's heaviest per-capita consumer of AI coding tools but a methodology importer. The export opportunity is teaching context engineering as literacy before the next foreign tool ships.

Terms used in this post

Agentic engineering — Karpathy's 2026 replacement for "vibe coding." Working with an AI agent the way you'd work with a teammate: define the task, review the diff, refuse the work if you can't explain what it does.

Context engineering — Designing the persistent information environment an AI agent operates inside: project rules, design docs, tools it can call, and feedback signals. The discipline that replaced prompt engineering as the skill that decides outcomes in 2026.

Harness engineering — The Toss-coined sibling of context engineering. Treats the workflow around the AI (the harness — skills, agents, hooks, state) as the main lever for org productivity, not the choice of model.

Match Rate — A bkit metric: the percentage of a shipped implementation that matches its design document. Below 90% triggers up to five automated improvement cycles before the user is asked to review.

RLS — Row-Level Security. A Supabase / Postgres feature that decides which rows a logged-in user can see. The most-broken setting in vibe-coded apps: writing auth.role() = 'authenticated' lets every logged-in user read every other user's rows, a silent breach that ships with the deploy.

Related reading:

• Vibe coding with context engineering
• Intent is the new craft
• bkit + harness engineering