All Gas Town, No Brakes Town

“Welcome to Gas Town” wrote Steve Yegge on the first day of the sextuple deuce, and obviously this is the kind of year 2026 is gonna be so let’s try to figure out what the hell that means. Yegge opens with:

This all made sense to me, because I am a horrible goblin with a soul defiled by the years I squandered questing after forbidden knowledge in the profane darkness of software Gehenna. But for those of you who still dwell in the Light:

• An IDE is an integrated development environment—basically an editor where you write and debug software.

• Claude Code is Anthropic’s software generating chatbot.

• Yak shaving (named for a holiday in Ren & Stimpy) is tedious and apparently purposeless activity that’s nevertheless necessary to achieve a goal. Programming is 98% yak shaving by volume.¹  

So Gas Town is a development environment for writing software using multiple code generating chatbots. It’s one answer to the question “what if we made the whole software company out of vibe coding?”

“Vibe coding” is the idea that nowadays you can just open a chat window and tell the AI to write your software for you without ever needing to read the code, and possibly without even knowing how to read it. Yegge co-wrote a book about vibe coding, and despite the term being less than a year old we’re already fighting about what it means. Can you imagine programmers arguing over semantics?²  

But in the year since Andrej Karpathy alleged that he could “see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works,” a lot of software developers have tried vibe coding and found that the most important word is “mostly.” Programming agents get stuck, neglect to perform essential steps, or just run out of context and forget what they were supposed to do. The most passionate vibe coding aficionados imagine infinite monkeys writing Shakespeare, but in practice it’s more like one semi-smart monkey copying and pasting half-understood example code until it kinda works, unless they get distracted first. If you’ve worked in software this might sound familiar, because it’s roughly the same experience as assigning tasks to a junior developer. Vibe coders are, in effect, attempting to move themselves up one level in the software worker hierarchy to either “senior developer” (if they’re reading the code) or “engineering manager” (if they‘re not).

Except for the real sickos, most people have always found writing code to be incredibly tedious. Programming languages themselves are an effort to move the job of instructing the computer from the most literal level, putting bits into registers, up into increasingly conceptual realms of describing “the way data flows through a system,” as Paul Ford wrote last week in a company newsletter about his own experiments with vibe coding. Like a Kübler-Ross for coders, Yegge described eight stages of succumbing to the lure of just letting the computer program its own damn self:

If vibe coders are trying to level themselves up to engineering manager, Gas Town is trying to skip its users all the way to CEO, with the entire management and workflow structure of a software development company recreated in AI and taking direct orders via a terminal multiplexer.

Steve Yegge is very intentional about naming and branding. This may seem surprising if you take a look at Clay Shirky‘s Gas Town glossary, which concisely defines the system’s various agents, roles, and workflows. “Gas Town” itself is a reference to the oil refinery citadel in Mad Max. The Mad Max theme continues in the names of the Mayor, the AI agent the user primarily communicates with; Rigs, which is what Gas Town calls projects under development; and the workflow concepts of the Refinery and the Convoy. You don’t care what those are. But the center cannot hold, and things start to fall apart when you get to the Dogs, named for the MI5 internal security characters in Mick Herron’s Slow Horses series of spy novels. The Dogs work for the Deacon, who Yegge says is “named for a Dennis Hopper character from Waterworld that was inspired by the character Lord Humungus in the Mad Max universe,” (??) but is also short for “daemon beacon” because its role is… well, I’ll let Yegge describe this too:

Remember that Mayor, Deacon, and Dogs are all AI agents. The existence of a whole subsystem dedicated to reminding each of them to do their jobs is a good glimpse into the current reliability of AI agents. There’s also a special upper-level Dog whose job is to remind the Deacon to do its job, which is to remind all the other workers to do their jobs. Lol. Quis custodiet, my dude?

Other elements of Gas Town have names drawn from Kurt Vonnegut’s novel Cat’s Cradle, the “Breaking Bad” TV show, and as far as I can tell some things are just named descriptively, like “the Crew.” The workflow is similarly a mishmash of concepts perhaps best exemplified by this detail from a slop-art diagram of the overall system:

You sling beads to a hook which activates a polecat according to GUPP. Jesse what the fuck are you talking about?

At first this all seems like gibberish, and it is. But I think Yegge is one of those people with an innate and preternatural sense of the power and purpose of naming things—someone who understands that names are marketing and marketing is not always about attracting the largest possible audience. In this case, the best outcome for Yegge is for Gas Town to appeal to a relatively small number of absolute sickos who vibe hard with his personal brand and who can usefully contribute to the project, and also for Gas Town to actively repel looky-loos and dilettantes like me (and probably you), who will only waste his time with a lot of stupid questions like “huh?” and “molecules?” and “did you say seances?” Oh yeah: there are seances. Don’t ask.

By this standard, Gas Town has apparently been very successful. It attracted enough attention that even I heard of it, in a crowded field of similar projects with names that make me want to literally die of boredom like “OpenAI AgentKit,” “Google Agent Development Kit,” and “Flowise.” These might not even be the same kind of thing—ironically the search results for “AI code orchestration” are so hopelessly polluted with AI-written slop that it’s difficult to tell. Either way, Gas Town is a chaotic conceptual mess that probably doesn’t work very well, but at least it has 🎷pizzazz✨.

It’s been several years since I wrote a line of code and inshallah I never will again so none of the above comes from firsthand experience. I still have friends and readers in the industry though, and my sense from them is that the worst job in software development right now is to be the person reviewing AI generated pull requests, which are often submitted by a junior developer who doesn’t know or care what the code actually does other than turn a failing test green. While Gas Town is still way out on the most gore-soaked of bleeding edges, AI is taking over programming much faster than I would have predicted it could. Is that a bad thing?

Personally I am an AI hater. Anthony Moser captured my attitude perfectly in his post from August titled “I Am An AI Hater:”

I’m unlikely to ever start using AI, in the same way that I’m unlikely to ever take up naval gunnery or political assassination. But like AI both of those things do exist, and while they’re always regrettable I can’t categorically say that I think either of them should never be used by anyone under any circumstances. The statistical machine learning and pattern-finding roots of generative software have a wide array of potentially useful applications, but even for generative software narrowly considered, writing code is one of the very few places where I could see it being useful as a tool.

Casey Newton recently used Claude to build a personal website, and reports that he had a good time doing it (and also that he still has a boyfriend). And as I mentioned above, no less a luminary than Paul Ford has been busy building a software consultancy organized around the premise that generative AI can make the prototyping and creation of business software orders of magnitude faster and cheaper. There are lots of industries that desperately need better software but won’t ever get it if they have to pay what it costs to write it by hand, which is incredibly expensive, so on the whole: big if true.

Clearly some people are having some success turning over the tedium of coding to the robots. But what kind of people? In his vibe coding post Paul wrote that a “shocking number of architectural decisions are not purely technical, but come down to taste…”

The irreplaceable role of taste is also apparent on Casey Newton’s AI-built website, with its page loading spinner, custom cursor, and pixel-block mouse follow effects. Casey is, notably, not a web designer.

But Steve Yegge has been a senior engineer at both Amazon and Google. Paul Ford has had a hand in more software projects than most people. One of the best coders I know is currently using AI for all of his work, and reports that it’s helpful and makes him more productive but requires constant and attentive supervision. All of these people are experts—they know deep in their bones how to describe the way a computer program should be structured and how its data should be managed. Their expertise was built through decades of hand writing code. And still, Paul’s semi-working synthesizer app has “taken 1,000 turns to get… to this state. How many turns remain? Another thousand? Twenty times that? I don’t know.”

Will the AI ever gain a high level conceptual understanding of how to structure software to be reliable and maintainable, when it isn’t currently capable of a high-level conceptual understanding of how to run a vending machine? Will today’s junior developers ever gain that understanding themselves if they spend their careers instructing the AI rather than writing code? Will there even be junior developers if all the senior devs are handing off that work to polecats or whatever?

Or in a larger sense, as Luke O’Neill put it in his book We Had It Coming, “I don’t understand where they think their personalized sense of aesthetics they plan to input into the computer is going to come from down the line when all human artists have been driven out of business.” No one knows. I guess we’ll find out!

I hope this has been useful to you, the Tabs reader, who has probably never heard of Gas Town and is not a software developer of any kind. Let it never be said that I know my audience or cater to your desires. This took longer than I wanted it to because writing informative essays is awful work—almost as bad as writing code. I really wish I'd listened to what my mother told me when I was young. Next Time: some actual Tabs probably.

1  The other 2% is bike-shedding, and I’m not even going to get into that because we’ll be here all day.

2  Do read that Simon Willison post. “What next? A book about prompt injection that’s actually about jailbreaking?” is one of the funniest lines I’ve ever read.