4 min readJust now

–

I’ve been thinking a lot about “vibe coding” or “agentic coding” and why everyone is having such different experiences.

On one end, you’ve got people on Twitter claiming they do basically 100% of their coding with agents / Claude Code / Codex / whatever. On the other end, you’ve got guys like The Primagen who can’t even get simple stuff to land reliably.

After trying to get close to “100% agentic” through December and January, my take is: vibe coding is just another hobby. And I don’t mean that as an insult.

If your goal is purely “ship the task,” then agents are useful, but you’ll hit a point where you go, “okay, this is dumb, I can finish this faster myself,” and you just do it. Especially when you’ve got tab autocomplete, you’re already in flow, code can be churned out pretty fast, and it’s code you probably understand more. That is the reality of using LLM’s. Sometimes the agent back-and-forth costs more than the work. It depends on the task, obviously. Some tasks are better left to machines, some for humans. The tradeoff is changing with Opus 4.5 and GPT 5.2.

But if you’re doing “100% vibe coding” then you’re accepting that tradeoff on purpose. You’re choosing a potentially slower path because the how is the point. Managing agents, crafting skills, building docs, designing workflows, wiring up the Factorio factory. That’s the fun part, the finished product is secondary.

The Primagen isn’t in the vibe coding hobby. His hobby is coding. His neovim setup, his keyboard, his keyboard layout, etc. He’s trying to use the thing like a normal tool. Input to output and done. If you approach it like that, you’re gonna get annoyed fast because it does fail in annoying ways, and you end up wrestling it for stuff you could’ve typed in three minutes. I don’t even know why I’m mentioning him that much, just the easiest public example. I probably could have just used myself as an example.

The 100% agentic crowd is in the hobby. They like building the car more than driving the car.

This is exactly like computer music.

If you just want to finish a track, Ableton or Logic or FL is faster. More instruments, better editing, better workflows. But there’s a whole subset of people who will choose an MPC or Maschine+ even though it’s slower and more constrained. Because it feels better. The workflow is satisfying. The song might end up basically the same to a listener but to the person making it the way they got there matters.

If you’ve ever been into Linux, you also understand this concept. Linux is useful, yes. But some people aren’t there because they need a working computer. They’re there because they wanted to install a bunch of distros, tweak their window manager, try out other window managers, organize their dotfiles, and get the system just right. The hobby isn’t the destination. The hobby is the path.

And yeah, you can run 10 agents. Or 20. It feels like you’re doing something huge. Sometimes you are. But a lot of the time, the context switching is brutal, and you’re not actually net faster. It’s basically the mythical man month but with bots.

What’s funny, the moment you hit a pain point, your instinct usually isn’t “stop.” Your instinct is “I should build a workflow to manage this automatically.” Which is the hobby, haha. That’s the tell.

Not saying agents aren’t powerful. They are. I’m doing stuff with them at work that we wouldn’t have budget or time for otherwise. Especially when you can let something churn in the background and come back to it. They’re great for bootstrapping, scaffolding, spiking, migrating, cleaning up a mess, drafting PRs, generating tests, writing docs, all of it.

But if the goal is “get 10 PRs out cleanly,” doing them one by one with you in the loop (with help from LLM) might just be faster and safer. I’ve found that maybe having 2 or 3 things in parallel, with some long running thing going on is kind of the sweetspot. Just depends on what I’m solving.

So yeah: the reason there’s such a huge discrepancy in opinions is that people are optimizing for different things.

• Some people optimize for output (just ship the task, whatever makes sense).
• Some people optimize for process (do it via agents).
• Some people want both, but the balance shifts from day to day.

And I think there’s this cultural thing under it, too. A lot of us got into programming at an early age because the programming itself is fun. If coding starts to feel commoditized, then agent orchestration becomes the new playground. It scratches the same itch. Tools, systems, workflows, leverage, cleverness. It replaces the craft feeling, just in a different form.

So if someone wants to get value from agentic coding without hating their life, the honest rule is this.

Treat it like something you’re going to be into, something you’re trying in good faith, not something you just use.

And this is really everything related to technology. It’s easier to do when you’re really in it. You’re into the language, you’re into the database, you’re into the framework, etc. Try learning something you’re forced to learn instead of something you want to learn, it’s a huge struggle.

If you’re not interested in building prompts, skills, docs, workflows, conventions, review loops, if you just want a tool, you’ll still get value, but you’ll probably hit a wall at “agent helps me start and accelerate parts of tasks.” The moment you force “100% agentic” without enjoying the process, it turns into frustration.