Skip to content
RuleSellBeta
Sign up
Back to blog

The Vibe Coding Arc: From Karpathy's Tweet to the May 2026 Successor

/RuleSell Team

Fifteen months from a single tweet to an industry-wide replacement term. Here is the month-by-month timeline of vibe coding's rise, peak, and quiet absorption into agentic engineering.

In February 2025, Andrej Karpathy posted a 280-character description of how he was writing code. By April 2026, the same Karpathy was on stage at Sequoia calling the same practice "passé" and introducing the successor term. In between, vibe coding became a Collins Dictionary Word of the Year contender, leaked 1.5 million API keys, spawned at least six unicorn-valued tools, and was quietly absorbed into a broader paradigm called agentic engineering.

The arc lasted fifteen months. The SERP is still six weeks behind it. Here's the complete timeline.

February 3, 2025: the tweet

Karpathy's original post is so short it fits in a footnote:

"There's a new kind of coding I call 'vibe coding', where you fully give in to the vibes, embrace exponentials, and forget that the code even exists."

He goes on to describe accepting Cursor's suggestions without reading them, asking the model to fix bugs by pasting error messages back, and "barely touching the keyboard." Simon Willison's annotation — published two days later — is the most-cited interpretation: vibe coding is a workflow for "throwaway weekend projects," not a software engineering methodology. That distinction would matter later.

The tweet was 26 million views by week-end. By month-end, "vibe coding" had a Wikipedia page.

March – May 2025: the tool wave

Tools that already existed got rebranded. Tools that didn't yet exist shipped fast. The vibe-coding tool stack of spring 2025 was:

  • Bolt.new — text-to-fullstack-app, Stackblitz's bet. Hit $40M ARR by July 2025.
  • Lovable — Swedish "vibe-code your SaaS" tool. Hit $50M ARR before its first birthday.
  • Replit Agent — the established player, repositioning hard.
  • Cursor Composer — the Pro-tier multi-file vibe loop.
  • v0 (Vercel) — for React/Next.js component generation.
  • Codeium / Windsurf — the IDE-native vibe assistant.
Each tool monetized the same loop: describe the app in prose, accept the first generation, iterate by chat. None of them had a name for what they were doing other than "AI coding," until Karpathy's term gave them a flag to plant.

June – August 2025: the consultancy bubble

By summer, vibe coding had a consulting class. Every AI consultancy added "vibe coding workshop" to their service menu. Y Combinator's Summer 2025 batch was 44% AI-first. Sequoia, a16z, and Greylock published think pieces. The phrase "vibe coding" got into the speaker bios at three different conferences.

This is also when the criticism started landing. Uncle Bob's blog post — "Vibe coding is not the same as disciplined agentic development" — was the first widely-shared dissent from a respected source. The HN response thread, "Vibe coding: what is it good for? Absolutely nothing", pulled 600+ comments. Even people who liked it conceded the failure modes.

September 2025: AGENTS.md and the discipline turn

September is the inflection. Three things shipped:

AGENTS.md launches. OpenAI's Codex CLI ships with a config file at the project root called AGENTS.md. It's tool-agnostic, OpenAI-led but published as an open standard, and by the end of the month it has 82 commits and 60,000+ projects using it (HN id 44957443, 382 comments). The point: developers are willing to write docs for agents that they wouldn't write for humans. This is the first sign that the industry wants more discipline, not less. Spec Kit ships. GitHub publishes a spec-first methodology for AI coding: "specify → plan → tasks → implement." The whole point is the opposite of vibe coding — formal spec before the model writes anything. Superpowers launches. Jesse Vincent (Prime Radiant) publishes a Claude Code methodology: brainstorm → plan → implement, with subagent-driven development, TDD discipline, and explicit verification. The HN original post and a rave review thread bring it into the mainstream. By April 2026 the repo is at 120k+ stars.

The discipline turn isn't coordinated, but it's unmistakable. The tools that ship in September 2025 all share an assumption: the agent works better when you tell it more, not less.

October – November 2025: the security wake-up

In October, Vibehouse leaks 1.5 million API keys. It's a vibe-coded admin panel with no auth, exposed publicly, scraped by a researcher who reported it before bad actors found it. The number gets cited everywhere. By December, Veracode's 2025 report says 45% of AI-generated code introduces security vulnerabilities.

The narrative shifts. Vibe coding is fine for a weekend prototype. Vibe coding is not fine for the thing that has your customers' payment data. Every blog post about vibe coding for the next quarter has a paragraph that starts "but for production…".

December 2025: Karpathy inverts

In a year-end thread, Karpathy admits his own ratio inverted. Where he used to write 80% of his own code and delegate 20% to agents, he now delegates 80% and writes 20%. The phrasing is casual but the implication is loud: even the term's inventor isn't doing pure vibe coding anymore. The replies fill with people saying the same thing happened to them in November.

January 2026: AGENTS.md becomes a standard

Linux Foundation announces the Agentic AI Foundation. AGENTS.md is one of the foundation's first stewardship projects. Tool adoption: Codex, Cursor, Copilot, Gemini, Jules, Amp, Factory all read AGENTS.md natively. The "vibe code your app" era was about removing structure. The AGENTS.md era is about putting it back, in a form the model can read.

Two of the three biggest LLM eval tools also get acquired in this window — Langfuse → ClickHouse in January, Promptfoo → OpenAI in March. Eval and observability mature in parallel to the discipline turn. (We covered the implications in a separate post.)

March 7, 2026: autoresearch

Karpathy publishes a new repo: autoresearch. The announcement tweet hits 8.6 million views in two days. The repo is 630 lines and explicitly says "you don't use it directly — you give it to your agent." Within six weeks, four Claude Code ports exist with combined 17k+ stars:

The pattern is the opposite of vibe coding. Vibe coding is give in to the vibes and ship. Autoresearch is define a measurable metric and iterate autonomously until it improves. The full pattern is in our autoresearch guide.

April 2026: Sequoia calls it

At Sequoia's AI Ascent 2026, Stephanie Zhan introduces Karpathy and uses the phrase that ends the era. Paraphrasing the conversation:

"Vibe coding raised the floor. Agentic engineering raises the ceiling."

Karpathy agrees, then is more direct: vibe coding is "passé." The replacement is agentic engineering: same AI tooling, more discipline, treating the agent like a junior developer who needs context, specs, evals, and review — not like a magic wand. Simon Willison's May 6 follow-up is the canonical synthesis: the same tools work for both, but you ship better software with the discipline.

May 2026: where we are

The vibe coding era is over in vocabulary terms — Karpathy moved on, Simon Willison moved on, the discourse moved on. But it's not over in practice. Bolt, Lovable, v0, Cursor Composer, and Replit Agent are all bigger than they were a year ago. The tools are the same; the methodology around them changed.

Here is the tool map for May 2026. Where each tool sits on the vibe ↔ discipline axis:

PhaseToolsWhat they reward
Pure vibe (Feb 2025)Cursor accept-all, Replit Agent, Bolt.new, Lovable, v0Speed-to-prototype
Vibe + Cursor rules (Spring 2025)Cursor with .cursorrules, Windsurf with .windsurfrulesSpeed + constraints
Methodology overlay (Sept 2025)Claude Code + CLAUDE.md, AGENTS.md, Spec Kit, SuperpowersRepeatable shipping
Discipline + autonomy (March 2026)Claude Code + skills + autoresearch + subagentsLong-horizon work
Agentic engineering (May 2026)Methodology + skills + evals + subagents + cross-tool configsProduction systems
The honest reading: the tool matters less than where you sit on the axis. A team using Cursor with disciplined rules + AGENTS.md + an eval pipeline is doing agentic engineering. A team using Claude Code with no CLAUDE.md, no skills, no review process is still vibe coding — it just has a more expensive model.

What replaced vibe coding (with specifics)

The "agentic engineering" successor is the umbrella term. Underneath it, three specific practices replaced the parts of vibe coding that broke.

1. Configuration files replaced ad-hoc prompts. CLAUDE.md, AGENTS.md, SKILL.md — three files that encode the project context the agent needs. The HumanLayer rule is that CLAUDE.md should be under 60 lines, because frontier models handle ~150-200 instructions reliably and lose discipline past that. The whole point is the opposite of vibe: tell the agent more, explicitly, ahead of time. 2. Skills replaced one-shot prompts. Anthropic's Claude Code skills system is the model-native version: a skill is a triggered context bundle that loads only when needed. The result is a 4-line context budget instead of a 400-line one. An always-on rule is a waste of context. A triggered skill is a precision instrument. RuleSell exists because the gap between "I have a skill" and "I have a skill that's been audited and won't break my codebase" is real. 3. Eval and review replaced "ship it and see." Both Hamel Husain and Shreya Shankar have been saying the same thing for two years: 60-80% of dev time on error analysis, review at least 100 traces, build the annotation UI into your app. The two eval-tool acquisitions in early 2026 (Langfuse, Promptfoo) made the tooling story messier but the principle is unchanged.

The five-step upgrade path

If you started in vibe coding and want to land in agentic engineering, here is the path that worked for the teams we've watched do it:

Step 1: Write a 50-line CLAUDE.md. Stack, conventions, deny-rules for what the agent should never touch. Done. Step 2: Convert your favorite repeat-prompts into skills. If you've typed "write a Next.js API route with Zod validation and Prisma" more than three times, that's a skill. See the build a Claude Code skill guide. Step 3: Add an AGENTS.md if your team uses more than one agent tool. The decision tree is short: one tool → CLAUDE.md only. Multiple tools → AGENTS.md primary, CLAUDE.md pointer. Step 4: Add evals for the part that ships to customers. Not for everything — for the customer-facing LLM features. DeepEval or Inspect AI are the open-source picks since Promptfoo got acquired. Step 5: Layer in autoresearch or subagents for long-horizon work. If you have a question that needs four hours of investigation, autoresearch is the pattern. If you have a task that has four parallel parts, subagents is the pattern. Don't add either before steps 1-4 are in place.

Where this analysis fails

We don't actually know if "agentic engineering" sticks as a term. Karpathy says it, Sequoia says it, Simon Willison says it. That's three points. The 2025 industry adoption of "vibe coding" had hundreds of points within weeks. The phrase could fade by August.

We also don't know which tools survive. Bolt, Lovable, and v0 are in different shapes than they were six months ago. Claude Code is gaining share. Cursor is defending. Aider is having a moment with the Polyglot benchmark. The substrate is moving faster than the methodology.

The thing we are confident about: the tools that get installed alongside discipline (eval, configs, skills, subagents) compound. The tools that get installed instead of discipline don't. That is the load-bearing finding from fifteen months of watching the arc.

What to read next

Sources