The argument in one line.
Vibe coding made building accessible to everyone, but agentic engineering — writing clear specs, supervising each step, and evaluating output before shipping — is what separates tools that scale from tools that break on the fifth user.
Read if. Skip if.
READ IF YOU ARE…
- You have used ChatGPT, Claude, or Lovable to build something and hit a wall when the project got serious.
- You want to turn AI skills into a service business or agency offer and need a framework that produces repeatable results.
- You already build with AI agents but have not formalized a review or evaluation loop before shipping.
- You are a non-developer who keeps getting fragile, one-off AI output and cannot figure out why it breaks.
SKIP IF…
- You are an experienced software engineer — Karpathy's original Sequoia talk is the primary source and has far more technical depth.
- You want hands-on implementation details for a specific tool; this video is conceptual framing, not a tutorial.
The full version, fast.
Karpathy's argument is simple: vibe coding raises the floor so anyone can build something, but agentic engineering raises the ceiling so professionals can build something reliable. His one-sentence definition — coordinating fallible agents while preserving correctness, security, taste, and maintainability — collapses into a single accountability claim: you are still responsible for your software. The practical translation is four behavioral shifts: replace vague prompts with written specs, replace blind trust with evaluation loops, replace one-shot attempts with supervised steps, and replace speed as the goal with quality. The host validates this not with theory but with a live case study: a GEO audit tool built on these principles earned 7,500 GitHub stars from developers, then generated a first paying client, and a community member replicated the client result from scratch following the same framework.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 01:14
01 · Vibe Coding Is Over
Hook invoking Karpathy's authority. Names the shift from vibe coding to agentic engineering and previews the video's structure.
01:14 – 01:45
02 · What Vibe Coding Means
Definition via relatable example: type a loose prompt into Lovable, get a landing page. No spec, no supervision. Screen recording of Vibecode UI shown.
01:45 – 02:18
03 · When Vibes Break
MenuGen case study: AI agent assumed Stripe email equals Google email, payment routed to wrong account. Tiny assumption, massive consequence.
02:18 – 03:23
04 · Agentic Engineering Explained
Floor/ceiling framing. Karpathy's exact definition quoted. The intern analogy. One-sentence thesis: you are still responsible for your software.
03:23 – 04:07
05 · Software 3.0 Shift
Three-era model with animated 1.0/2.0/3.0 transitions. LLM as computer, natural language as interface, agents as runtime.
04:07 – 05:39
06 · Four Key Shifts
Numbered behavioral changes. Shift 1: specs over vague prompts. Shift 2: evaluation loops over blind trust. Shift 3: supervised steps over one-shots. Shift 4: quality over speed.
05:39 – 06:41
07 · GEO Tool Case Study
Creator's own agentic-built GEO audit tool. Detailed specs, coordinated sub-agents, quality bars, PDF output. Result: 7,500 GitHub stars from developers without marketing.
06:41 – 07:51
08 · From Stars To Clients
Real audits sent to business owners. Agency-quality PDF reports. First paying client. Community member replication from scratch, same outcome.
07:51 – 08:39
09 · How To Start Today
Three-step loop: pick one project, write spec like handing to an intern, break into supervised steps, evaluate before shipping. THE SPEC orbital diagram shown.
08:39 – 09:22
10 · Wrap Up And CTA
Restates the line in the sand. Community and tool links. Subscribe ask with channel shelf shown.
Lines worth screenshotting.
- Vibe coding raises the floor; agentic engineering raises the ceiling — they are not competing approaches, they are different altitude targets.
- The single most dangerous assumption in vibe coding is that the AI will catch its own mistakes before your client does.
- Karpathy's MenuGen bug — AI assumed Stripe email equals Google email, payment went to the wrong account — is a real product failure from a Sequoia-level builder.
- Writing the spec like you are handing it to an intern who has never met you is the entire discipline of agentic engineering in one instruction.
- Seven thousand five hundred GitHub stars from developers who starred because the architecture was clean is stronger social proof than any marketing campaign.
- The people who win in AI from here forward are not the fastest prompters — they are the best agent architects.
- Software 3.0: the LLM is the computer, natural language is the programming interface, agents are the runtime.
- AI gives you a team of interns. If you do not give them clear instructions and supervise their work, they will ship something embarrassing with your name on it.
- A community member learning from scratch followed the same agentic principles and landed a paying client in weeks — the result is repeatable, not luck.
- The real test of an AI-built output is not whether it looks right; it is whether you would put your name on it and send it to a paying client.
- Five seconds of checking before you ship — does this number make sense, does this claim match reality, did it hallucinate anything — is the entire evaluation loop.
- GEO (generative engine optimization) is SEO for AI search engines: ChatGPT, Perplexity, Google AI overviews — and it is a service almost no agency is selling yet.
Four habits that separate reliable AI output from fragile demos.
WHAT TO LEARN
The difference between an AI tool that embarrasses you in front of a client and one that earns 7,500 GitHub stars comes down to four behavioral shifts, not four different tools.
- A vague prompt produces a vague result every time — replacing 'make me a marketing plan' with a five-page spec that includes audience personas, three ad headlines, and a 90-day calendar is not extra work, it is the minimum viable instruction.
- The MenuGen bug — AI assumed Stripe email equals Google email, payment went to the wrong place — happened because nobody built an evaluation loop; five seconds of checking before shipping prevents the class of errors that end client relationships.
- Complex projects fail when handed to AI all at once; breaking a five-step project into step-review-step-review cycles is not slower, it is how you catch the hallucinated logic before it reaches production.
- Speed was the selling point of vibe coding, and speed is also why vibe-coded tools break for the fifth user; the next era rewards builders who treat quality as the constraint, not the afterthought.
- The GEO audit tool case study proves the framework is repeatable: the creator built it with detailed specs and coordinated agents, a community member with no prior experience followed the same principles, and both landed a paying client within weeks.
- Writing a spec 'like you are handing it to an intern who has never met you' — defining success, output format, what can go wrong, and what cannot go wrong — is the entire discipline of agentic engineering collapsed into a single instruction.
Terms worth knowing.
- Vibe coding
- The practice of prompting an AI tool with a loose natural-language description and accepting whatever it generates without writing formal specifications, testing, or supervision. Produces fast first versions but fragile production software.
- Agentic engineering
- Karpathy's term for coordinating AI agents with clear specs, structured review loops, and quality gates — treating the AI as a team of interns you supervise rather than an oracle you trust blindly.
- Software 3.0
- Karpathy's label for the current era: the LLM itself is the computer, natural language is the programming interface, and agents are the runtime. Follows 1.0 (hand-written code) and 2.0 (neural networks trained on data).
- GEO (Generative Engine Optimization)
- The practice of optimizing content and structured data so that AI search engines — ChatGPT, Perplexity, Google AI Overviews — surface and cite your material. The AI-era equivalent of traditional SEO.
- Evaluation loop
- A deliberate pause between receiving AI output and shipping it, during which you verify the numbers, check for hallucinations, and confirm the output matches reality. The minimum viable version takes five seconds.
Things they pointed at.
Lines you could clip.
02:22
“Vibe coding raises the floor. Agentic engineering raises the ceiling.”
02:57
“You're still responsible for your software. That's the whole talk in one sentence.”
03:06
“The AI doesn't replace you, it gives you a team of interns. And like any team of interns, if you don't give them clear instructions and supervise their work, they're going to ship something embarrassing with your name on it.”
03:49
“We're now literally programming in English. Not as a toy, not as a hobby project, as the actual way to build software going forward.”
07:38
“The people winning in AI from here forward aren't going to be ones who can prompt the fastest. They're going to be the ones who can architect AI agents the best.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogystory
00:00Vibe coding is dead or at least the version of it we have all gotten used to. Because for the past year, everyone has been vibe coding. Developers, startup founders, AI creators, everyone on Twitter.
00:11You'd open Clog, ChatGPT, Cursor, or Lovable, type type in an idea and suddenly you have an app, a dashboard, a landing page, maybe you're in a full product. But what if I told you that Andre Karpathy, the person who helped make vibe coding popular in the first place is now saying that the era of vibe coding is over.
00:29Not me, him. The guy who co founded OpenAI, ran AI at Tesla, and now works at Entropic, the company behind Claude, is saying we're moving into a new phase called agentic engineering. And this matters because the future is not just prompting AI to write code, It's learning how to work with AI systems that can plan, use tools, check their own work, fix mistakes, follow constraints, and help you build real software and automations that actually work beyond a demo.
00:58So in this video, I'm going to break down exactly what Karpathy said, what agentic engineering actually means, why vibe coding by itself is not enough anymore, and why I think this changes everything. Especially if you're using tools like cloth code, MCPs, or AI agents to build real systems for businesses.
01:15Okay. So first, let's talk about what vibe coding actually is. Because if you've used ChatGPT or Clot or Lovable to build anything in the past year, you've probably done it.
01:23Vibe coding is when you type something like build me a landing page that sells protein powder to runners. Right? And AI just figures it out.
01:31No spec, no requirements, no supervision. You just vibe with it essentially.
01:35And for a year, this felt like magic. Anyone could build anything. You didn't need to understand what was happening under the hood.
01:42You just had to describe what you wanted. And AI took care of the rest. But here's what nobody was talking about.
01:47The output was sketchy, bloated code, hallucinated logic, security holes nobody caught, tools that would work for the first user and broke for the fifth. Karpathy actually shared a brutal example.
01:58He built a real product called MenuGen. The AI agent he was using accidentally matched a customer's Stripe email to their Google email.
02:07Just assume that they will be the same. They weren't. The payment went to the wrong place.
02:11Tiny detail, massive consequences. That's vibe coding in a nutshell.
02:16It looks like it works until it doesn't. So what does Karpathy say we should do instead? Agentic engineering.
02:22And here's the line that stopped me cold. Vibe coating . Agentic engineering .
02:28Now let me break that down. Because vibe coating raises the floor. This means that anyone can now build something.
02:35That's a democratization. But agentic engineering raises the ceiling. Meaning professionals can now build something really good, reliable, secure, maintainable.
02:45Tools that don't fall apart the moment your fifth customer signs up. His exact definition was and I'm quoting here. Coordinating fallible agents while preserving correctness, security, taste, and maintainability.
02:57And the line he kept coming back to throughout the entire talk was this. You're still responsible for your software. That's the whole talk in one sentence.
03:06The AI doesn't replace you, it gives you a team of interns. And like any team of interns, if you don't give them clear instructions and supervise their work, they're going to ship something embarrassing with your name on it. Karpathy actually broke this whole thing down in three eras of software.
03:22And once you see it, you can't unsee it. Software one point o, which is humans wrote code by hand. Decades of it.
03:29Every line of every program manually typed. Software two point o, it was neural networks trained on data. Instead of writing rules, we train machines to figure out the rules themselves.
03:40And now, software three point o. This is where it gets crazy. The LLM itself becomes the computer.
03:47Natural language becomes the programming interface. And agents become the runtime. Read that again.
03:52We're now literally programming in English. Not as a toy, not as a hobby project, as the actual way to build software going forward. That's the huge shift.
04:02That's what Karpathy was telling Silicon Valley two weeks ago. And that's why everything is about to change. Now here's the part that most creators covering this talk are missing.
04:10Karpathy was speaking directly to developers, but the principles he laid out apply to anyone building anything with AI, even if you have never written a single line of code in your life. I'm going to give you four shifts you need to make starting today. One, stop vague prompting.
04:26Start writing clear specs. Don't say make me a marketing plan. Say produce a five page marketing plan for a b to b SaaS company targeting health care clinics.
04:36Include audience personas, three ad headlines, and a ninety day content calendar.
04:42The more specific the spec, the better the output. Every single time. Two, stop trusting AI output blindly.
04:49Build evaluation loops. Before you ship anything that AI gives you, ask yourself, does this number make sense? Does this claim match reality?
04:57Did it hallucinate anything? Those five seconds of checking save you from looking like a fool in front of the real client. Three, stop trying to one shot complex projects.
05:07Break them into supervised steps. If your project needs five things, don't ask AI to do all five of them at once. Step one, review.
05:15Step two, review. That's literally what coordinating fallible agent means. Four, stop chasing speed.
05:22Start chasing quality. Speed is what got us vibe coding in the first place. Quality is what going to define the next era.
05:29That's the framework. And here's what's wild. When I started looking at my own work through this lens, I realized I've been applying these exact principles for months now.
05:38Let me show you what I mean. About four months ago, I built something called a GEO audit tool. Now GEO stands for generative engine optimization.
05:46Basically, SEO for AI search engines like ChatGPT, Perplexity, and Google AI overviews. I didn't vibe code it.
05:53I wrote detailed specifications. I broke it into coordinated sub agents. I set quality bars and scoring methodologies.
05:59I built it to produce structured professional PDF reports. Then I pushed it to GitHub.
06:05And now it has 7,500 stars from developers from all over the world.
06:11Now think about that for a second. Developers, the exact audience Karpathy was speaking to at Sequoia looked at what I built and started. Now, they didn't start it because of marketing.
06:21They didn't start it because of social media. They started because the architecture was clean. The agents had defined roles.
06:27The quality bar was real. And here's the thing. I didn't know I was doing agentic engineering.
06:32I was just trying to build something that actually worked. But that's the entire point. When you apply the principles even without knowing the name for them, the output speaks for itself.
06:41But the real test isn't GitHub Stars. GitHub Stars don't pay bills. The real test is whether the output is good enough that someone will pay for it.
06:49So my team took that same tool and used it to run real audits for real businesses. We analyze a website, generate the PDF report, and send it over to the business owner. The output looks something like an SEO agency would charge thousands of dollars to produce.
07:03Real data, real audits, specific recommendations with priority and impact ratings.
07:09My team landed actually our first paying client within a few weeks. And here's the part that actually convinced me this wasn't just luck. Someone in my community, someone learning all of this from scratch took the same approach.
07:21Built her version of the tool, followed the same principles, and landed her first paying client within few weeks. Same approach, different person. Same outcome.
07:30That's when it clicked for me. This isn't about being a developer. It's not about being a marketer.
07:35It's about applying these principles correctly. And when you do, the results are repeatable. That's exactly what Karpathy was pointing at.
07:42The people winning in AI from here forward aren't going to be ones who can prompt the fastest. They're going to be the ones who can architect AI agents the best. So then, how do you actually start doing this in your own work?
07:54Pick one project. Don't try to build an empire. One specific problem you want to solve with AI.
08:00Write the spec like you're handing it to an intern who's never met you. What does success look like? What's the output format?
08:07What can go wrong? What can't go wrong? Then break it into pieces.
08:10Don't ask AI to do the whole thing at once. Step one, supervise. Step two, supervise.
08:17Step three, and before you ship anything, evaluate. Does it match reality? Would you put your name on it?
08:22Would you send it to a paying client? That's it. That's agentic engineering applied to real life.
08:28Whether you're building tools, writing content, doing research, or trying to start an agency, the framework works. The proof is out there.
08:37The only question is whether you're going to apply it. If you want to see exactly what this looks like in practice, I've got every tool I've built using these principles available plus a community where we go deep on how to actually do this end to end. Link is going to be in the description.
08:51If you want to check it out, no pressure either way. Now, Karpathy didn't just give a developer talk at Sequoia two weeks ago. He drew a line in the sand.
08:59Vibe coding is the past. Agentic engineering is the future. And the people who win in AI from here on out aren't the ones who can type the fastest.
09:07They're the ones who can direct AI agents the best. Okay. So you've got the framework now.
09:12Go ahead and build something. And if this clicked for you, hit subscribe. I'm going to keep breaking down exactly what this looks like from practice in this channel.
09:20So hopefully, I'll see you in the next one.
The Hook
The bait, then the rug-pull.
The title borrows Andrej Karpathy's name to stop the scroll, then the first sentence earns the click: vibe coding is not just fading, the specific version everyone normalized over the past year is over. What follows is a clean translation of Karpathy's Sequoia talk for anyone who builds with AI but has never touched a compiler.
Named ideas worth stealing.
02:22concept
Floor/Ceiling Frame
Vibe coding raises the floor — democratizes building so anyone can create something. Agentic engineering raises the ceiling — enables professionals to build something reliable, secure, and maintainable.
Steal forany positioning deck that explains why your AI service is different from a chatbot
03:23model
Karpathy's Three Eras of Software
- 1.0: humans write code by hand
- 2.0: neural networks trained on data
- 3.0: LLM is the computer, natural language is the interface, agents are the runtime
Historical frame that makes Software 3.0 feel inevitable rather than hype.
Steal forany talk or video explaining why now is the inflection point for AI builders
04:07list
The Four Shifts
- Stop vague prompting — write clear specs
- Stop trusting AI output blindly — build evaluation loops
- Stop one-shotting complex projects — break into supervised steps
- Stop chasing speed — start chasing quality
Behavioral translation of agentic engineering for non-developers. Each shift is framed as a STOP/START pair.
Steal fornewsletter issues, Twitter threads, or workshop curriculum on AI workflow discipline
08:10model
The Spec Orbit
- What does success look like?
- What is the output format?
- What can go wrong?
- What cannot go wrong?
Four questions that orbit every well-written spec. Visualized as a planetary diagram in the video.
Steal forany onboarding doc or client brief template for AI-assisted work
How they asked for the click.
VERBAL ASK
08:39link
“I've got every tool I've built using these principles available plus a community where we go deep on how to actually do this end to end. Link is going to be in the description.”
Soft and low-pressure ('no pressure either way'). Community-forward rather than product pitch. Discovery call mentioned only in description, not verbally.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
