The Emergent Podcast — Episode 9Vibe Coding to Agentic Engineering: When Everyone Can Build, What Matters Is What You Build"AI is now awake. And it's a big contrast to even two, three months ago." — Nick Baguley

Listen on: Apple Podcasts · Spotify · YouTube · RSS

Episode Duration: ~1 hr 40 min | Published: 2026 | Season 1, Episode 9

🎙️ Episode Summary

One tweet changed a word. The word changed an industry. The industry is changing what it means to build.

In February 2025, Andrej Karpathy — co-founder of OpenAI and former head of AI at Tesla — published a single post coining the term "vibe coding": describe what you want in plain English, accept all AI-generated code without reading the diffs, and just… vibe. Twelve months later, it became the Collins Dictionary Word of the Year, 92% of U.S. developers use AI coding tools daily, 41% of all code is AI-generated — and Karpathy himself has already declared it passé, rebranding the practice as "agentic engineering."

In Episode 9, Justin Harnish and Nick Baguley dig into what really happened in that extraordinary year. Both hosts share their personal workflows and real projects — including Justin's intermittent fasting app, his vision of a personal "digital brain" with AI-queryable embeddings, and Nick's AI-native marketplace designed for both human and agent users. They navigate the empirical gut-punch of the METR study(developers are actually 19% slower on mature codebases using AI), the existential labor market questions (traditional programmer roles down 27.5% since ChatGPT's launch), and the philosophical territory that has been the Emergent Podcast's throughline since Episode 1: when code becomes a commodity, what becomes scarce?

Their answer: responsible agency — the judgment to decide what should be built, for whom, and with what values. That, they argue, is the skill that neither automation nor benchmarks can yet replicate.

📚 Resources & Reading List

Every link mentioned or referenced in this episode. Organized by theme for your exploration.

🔑 The Origin & The Debate (Required Reading)

1. Andrej Karpathy's Original "Vibe Coding" Tweet (Feb 2, 2025)
2. The tweet that launched the year. Karpathy describes accepting all AI code without reading diffs, pasting errors back without comment, and letting the codebase grow beyond comprehension. Note the caveat he included that industry largely ignored: "not too bad for throwaway weekend projects."
3. Karpathy's 2025 LLM Year in Review — bearblog.dev
4. His retrospective on vibe coding's arc from shower-thought tweet to Collins Dictionary Word of the Year. Key insight: "Code is suddenly free, ephemeral, malleable, discardable after single use." He also identifies Claude Code as the first convincing LLM agent.
5. Karpathy on "Agentic Engineering" (Feb 2026) — The New Stack
6. One year after coining vibe coding, Karpathy declares it passé. His new frame — agentic engineering— emphasizes that professionals orchestrate AI agents 99% of the time, with zero compromise on software quality. The rebrand is the narrative bookend of this episode.
7. Simon Willison — "Not All AI-Assisted Programming Is Vibe Coding" (Mar 2025) — simonwillison.net
8. The essential distinction: "If an LLM wrote every line of your code, but you've reviewed, tested, and understood it all, that's not vibe coding — that's using an LLM as a very fast typist." Also contains Willison's generous vision: "Everyone deserves the ability to automate tedious tasks."
9. METR Study: AI Makes Experienced Devs 19% Slower (Jul 2025) — metr.org
10. The empirical gut-punch of the episode. 16 experienced open-source developers, 246 real-world tasks. They believed AI made them 20% faster; they were actually 19% slower on their own mature codebases. Full paper: arxiv.org/abs/2507.09089
11. Vibe Coding — Wikipedia
12. Surprisingly rigorous. Tracks the full timeline, Lovable's 170 vulnerable apps, CodeRabbit's finding that AI code has 1.7× more major issues, Y Combinator stats (25% of W25 startups are 95% AI-coded), and the "vibe coding hangover" reported by Fast Company.

📖 Supplemental: The Deeper Cuts

• Scott H. Young — "Is Vibe Coding the Future of Skilled Work?"
• The variance argument: vibe coding may make software both much worse and much better simultaneously. Also argues that conceptual knowledge becomes more, not less, important when AI writes the code. A crucial counterweight to pure optimism.
• IBM — "What Is Vibe Coding?"
• Enterprise-oriented overview. Useful on the agile alignment: vibe coding fits fast-prototyping and iterative development. Contains the key qualifier Nick and Justin both echo: "AI generates code, but creativity, goal alignment, and out-of-the-box thinking remain uniquely human."
• Google Cloud — "Vibe Coding Explained: Tools and Guides"
• Practical tool comparison from Google's perspective — AI Studio, Firebase Studio, Gemini Code Assist. Useful for understanding which tool fits which use case.
• Software Engineering Job Market Outlook for 2026 — Final Round AI
• Data from Indeed/FRED and BLS projections. The key line: "In 2026, simply learning how to write code won't be enough. What really matters is understanding how code works."
• Top Vibe Coding Statistics & Trends [2026] — Second Talent
• The stat goldmine: 92% of US devs use AI daily, 41% of code is AI-generated, 74% report increased productivity, 63% of vibe coding users are non-developers, $4.7B market projected to reach $12.3B by 2027.
• How AI Vibe Coding Is Destroying Junior Developers' Careers — Final Round AI
• The counterpoint to the democratization narrative. Software dev job openings down 70%. The "new tutorial hell": learning without learning.
• Best AI Code Editor: Cursor vs Windsurf vs Replit — AIMultiple
• Head-to-head benchmarks of Claude Code, Cline, Cursor, Windsurf, and Replit Agent across API development and app-building tasks.
• 10 Claude Code Alternatives for AI-Powered Coding — DigitalOcean
• Solid comparison of the full 2026 AI coding landscape: Claude Code, Gemini CLI, Cursor, Replit, Windsurf, GitHub Copilot, Aider, and more.

📘 Books Referenced

• David Chalmers — Reality+: Virtual Worlds and the Philosophy of Mind — Justin's reference point for the holographic/digital substrate of reality; the "redness of red" and the hard problem of consciousness.
• 🔗 Publisher page
• Brian Christian — The Alignment Problem (revisited from Episode 4) — When code writes itself, alignment between human intent and machine output becomes the core individual skill, not just a civilizational concern.
• 🔗 brianchristian.org
• Eliezer Yudkowsky — "If Anybody Builds It, Everybody Dies" — Referenced in the consciousness/alignment close: the parable of the alien observer and the selfish gene's 200,000-year objective function vs. human contraception and saccharin.

🎙️ Creators & Thinkers Mentioned

• Andrej Karpathy — Co-founder of OpenAI, former Tesla AI head, coined "vibe coding," now advocating "agentic engineering"
• Simon Willison — Django co-creator; the clearest thinker on the vibe coding/AI-assisted programming distinction
• Nate B. Jones — Former head of Amazon product; YouTube + Substack on AI's labor market implications. Justin credits him for shifting his own optimism.
• Demis Hassabis — DeepMind CEO, AlphaFold creator, Nobel laureate in chemistry: "First we solve intelligence, then we solve everything else."
• Ray Kurzweil — Singularity theorist; the accelerating model capability doubling time (now ~7 months) maps his predictions.
• Eliezer Yudkowsky — AI safety researcher; the "selfish gene vs. consciousness" parable used in the closing alignment argument.
• David Chalmers — Philosopher of mind; the hard problem and Mary's Room as frameworks for why alignment requires more than an objective function.

💡 Key Ideas From This Episode

Concepts worth carrying into your week:

The Three Stages of AI Coding Consciousness (Nick's framework)

LLMs hallucinating → deep REM dream (GPT-3.5 era) → lucid dreaming (vibe coding, 2025) → fully awake (agentic engineering, 2026). The metaphor does real work: it explains why the same underlying technology feels categorically different at each stage.

"Responsible Agency" as the New Scarce Resource (Nick's closing argument)

When everyone can generate code, video, audio, and content, what can't be automated is the choice of what to build, for whom, and to what standard of taste. Judgment, systems thinking, and the willingness to exercise agency — these are the non-fungible skills.

The PRD as Demo (Both hosts)

A product requirement document is no longer a written specification — it's a working prototype. "The PRD today should be a full-blown app. Here's my demo; this is what acceptance criteria looks like. Now go make this production." The vibe-coded demo becomes the spec.

The METR Paradox

Developers believe AI makes them ~20% faster. Empirically, they are 19% slower on mature codebases. Possible causes: context-switching overhead, review burden, the seductive illusion of speed when tokens flow fast. The lesson isn't "AI doesn't help" — it's that measurement must catch up to method.

"The Experience Is the Point" (Justin's closing)

Even as models approach inductive reasoning and potentially displace the need for syntax-literate humans, Justin argues consciousness — the felt quality of experience — remains irreducibly important for alignment. Mary in the black-and-white room knows everything about color and still learns something when she sees red for the first time. That remainder is what makes alignment a hard problem, not just a technical one.

Sonnet 4.6 as "Staff Engineer" (Nick)

GPT-4 era → junior developer. GPT-5 era → mid-level. Claude Sonnet 4.6 + the right tooling → staff/principal engineer. With agentic harnesses, you're now talking about an engineering organization, not an assistant.

🔥 Quotable Moments"I don't code. I've taken coding classes. I've got a technical degree in chemical engineering. Fast forward to vibe coding: I'm losing sleep over not being in front of a computer."— Justin Harnish"It feels a little bit like spending your life trying to become a bodybuilder, and then you show up for the competition and realize the job is to push feathers around."— Nick Baguley"Claude Code was written in two weeks by four engineers. 90% of it was written by Anthropic agents working on that codebase."— Justin Harnish"When everybody can generate code, when they can generate videos and images and audio — the real scarce resource becomes responsible agency."— Nick Baguley"The universe deserves to be experienced. It is the best part of it. Even with all of this fun — the fact that it is like something to be in this life is the best part."— Justin Harnish"A markdown file shot a $220 billion hole — the SaaS apocalypse — into the legal research and much of the rest of SaaS."— Justin Harnish"If I could go back two years ago and have access to the tools I use today, I could do what a thousand engineers were doing at the time. It's like taking an iPhone back to the 1800s."— Nick BaguleySubscribe: Apple Podcasts · Spotify · YouTube · RSS

Contact: justinaharnish.com

More from Justin:

Substack: OrdinaryIlluminated.com

YouTube: Notes from the Vault - youtube.com/@justinaharnish

Web: Justinaharnish.com

Research: consciousgpt.org

The Emergent Podcast explores the Age of Inflection in Intelligence — tracing how new systems of thought, technology, economics, and culture emerge from the moment we are living through. New episodes released regularly.

© The Emergent Podcast | justinaharnish.com