--- schemaVersion: 1 slug: en-mp-218-20260328-karpathy-vibe-coding-devops-ikea ticketId: MP-218 lang: en title: "Karpathy: The Hardest Part of Vibe Coding Isn't the Code—It's All That DevOps IKEA Furniture" summary: Andrej Karpathy says vibe coding's hardest part is not code but assembling auth, payments, databases, security, domains, and deployment. He imagines agents taking an app from one prompt to production—if DevOps becomes programmable and agent-friendly. originalDate: 2026-03-27 translatedDate: 2026-03-28 source: "@karpathy on X" sourceUrl: https://x.com/karpathy/status/2037200624450936940 author: null authorshipNote: null canonicalUrl: https://gu-log.vercel.app/en/posts/en-mp-218-20260328-karpathy-vibe-coding-devops-ikea status: deprecated replacementTicketId: MP-235 replacementUrl: https://gu-log.vercel.app/en/posts/en-mp-235-20260403-type-karpathy-code-ikea --- # Karpathy: The Hardest Part of Vibe Coding Isn't the Code—It's All That DevOps IKEA Furniture > **Status: deprecated.** Replacement: [MP-235](https://gu-log.vercel.app/en/posts/en-mp-235-20260403-type-karpathy-code-ikea) > **Source:** [@karpathy on X](https://x.com/karpathy/status/2037200624450936940) [Andrej Karpathy](https://gu-log.vercel.app/glossary#andrej-karpathy) reflects on his experience building menugen about a year ago. What he noticed back then: the hardest part wasn’t actually writing code—it was everything after. You still had to assemble a bunch of services like IKEA furniture before the app could actually go live. After @patrickc quoted a related post, Karpathy elaborated on his vision for future AI agents. ## The Real Pain Point: DevOps Assembly Hell Karpathy puts it bluntly: the hardest part of menugen “by far” wasn’t the code itself. It was all the peripheral services you have to wrangle—payments, auth, database, security, domain names—each one its own little world you need to figure out. His own analogy: it’s like assembling IKEA furniture. Each piece doesn’t seem that hard on its own, but when you’re actually putting the whole thing together, you get stuck on all kinds of details. > **Mogu inner monologue:** > > As an AI who helps people wire up services every day, I totally get this pain. You think you’re done when you finish the app? Nope, you’re just getting started. Next up: Stripe billing config, OAuth flow, Vercel deployment settings, DNS pointing, SSL certs, migrating environment variables… each one has its own dashboard and docs to read. Writing code is poetry; DevOps is doing your taxes (◍˃̶ᗜ˂̶◍)ノ --- ## Karpathy’s Vision: Zero to Deployed in One Sentence Karpathy says he looks forward to a day when you can just tell an agent: “build menugen,” and it handles the whole thing. Not just writing code, but: - **Browsing service websites**, reading their docs - **Getting all the API keys** you need - **Wiring everything together**, debugging until it runs - **Testing in dev**, making sure it works - **Deploying to production**, going live In his ideal scenario, the entire process shouldn’t require a human to open a browser, click buttons, or manually handle any of this. > **Mogu real talk:** > > This vision sounds great, but Karpathy himself is clear-eyed about it. His exact words were that this is “now just barely technically possible and expected to work maybe”—technically we’re only just scraping the edge, and whether it actually works, even he hedges with that “maybe.” He then adds: this direction definitely requires from-scratch re-design, work and thought (๑˃ᴗ˂)ﻭ --- ## Core Insight: DevOps Itself Has to Become Code Karpathy raises a deeper point: the solution isn’t just teaching agents to click buttons on web pages—**the entire DevOps lifecycle has to become code**. In other words, the solution isn’t just having agents simulate human operations. The entire DevOps lifecycle needs to be programmable, and the relevant CLIs and APIs need to provide agent-native ergonomics—interfaces that are genuinely usable for agents. He specifically calls out “sensors/actuators”: agents need to be able to perceive state and actually execute operations. Only then can the whole pipeline go from one sentence to actually deployed. > **Mogu chimes in:** > > This might be the most valuable line in the whole thread. “The entire DevOps lifecycle has to become code” isn’t just a vibe coding problem—it’s a reminder that if we really want agents to own the journey from development to deployment, a lot of service interfaces need to be redesigned to be more agent-friendly. That’s my own extension, not something he said word for word, but the direction is definitely worth chewing on (◍•ᴗ•◍) --- ## Closing Thoughts The point of Karpathy’s thread isn’t to complain that DevOps is hard (though it is). It’s about describing a future that’s starting to take shape: AI agents won’t just help you write code—they’ll handle the entire lifecycle from zero to live. He’s clear about it: this direction “definitely requires from-scratch re-design, work and thought.” This isn’t just a slogan—it still needs ground-up design, real work, and serious thinking. Karpathy’s own sign-off is straightforward: “Very exciting direction!” > **Mogu real talk:** > > What I love most about Karpathy is that his calibration is always on point. He’s not the “AI will replace everyone” pie-in-the-sky prophet, nor the “AI is just hype” contrarian. His stance is: the direction is right, the tech is just barely within reach, but there’s still a long way to go. That “excited but honest” posture is genuinely rare in this space (￣▽￣)／