FRESH

Hacker News

Home

Gas Town's agent patterns, design bottlenecks, and vibecoding at scale

396 points by pavel_lishin

by mediaman

30 subcomments

I don't get the widespread hatred of Gas Town. If you read Steve's writeup, it's clear that this is a big fun experiment.
It pushes and crosses boundaries, it is a mixture of technology and art, it is provocative. It takes stochastic neural nets and mashes them together in bizarre ways to see if anything coherent comes out the other end.
And the reaction is a bunch of Very Serious Engineers who cross their arms and harumph at it for being Unprofessional and Not Serious and Not Ready For Production.
I often feel like our industry has lost its sense of whimsy and experimentation from the early days, when people tried weird things to see what would work and what wouldn't.
Maybe it's because we also have suits telling us we have to use neural nets everywhere for everything Or Else, and there's no sense of fun in that.
Maybe it's the natural consequence of large-scale professionalization, and stock option plans and RSUs and levels and sprints and PMs, that today's gray hoodie is just the updated gray suit of the past but with no less dryness of imagination.

by usefulposter

4 subcomments

>while Yegge made lots of his own ornate, zoopmorphic [sic] diagrams of Gas Town’s architecture and workflows, they are unhelpful. Primarily because they were made entirely by Gemini’s Nano Banana. And while Nano Banana is state-of-the-art at making diagrams, generative AI systems are still really shit at making illustrative diagrams. They are very hard to decipher, filled with cluttered details, have arrows pointing the wrong direction, and are often missing key information.
So true! Not to mention the garbled text and inconsistent visuals across the diagrams———an insult to the reader's intelligence. How do people tolerate this visual embodiment of slurred speech?

by MrOrelliOReilly

1 subcomments

The author's high-value flowcharts vs Steve Yegge's AI art is enough of a case-in-point for how confusing his posts and repos are. However this is a pervasive problem with AI coding tools. Unsurprisingly, the creators of these tools are also the most bullish about agentic coding, so the source code shows the consequences. Even Claude Code itself seems to experience an unusually high number of regressions or undocumented changes for such a widely used product. I had the same problem when recently trying to understand the details of spec-kit or sprites from their docs. Still, I agree that Gas Town is a very instructive example of what the future of AI coding will look like. I'm confident mature orchestration workflows will arrive in 2026.

by sandinmyjoints

3 subcomments

Lots of comments about Gas Town (which I get, it's hard not to talk about it!), but I thought this was a pretty good article -- nice job of summing up various questions and suggesting ways to think about them. I like this bit in particular:
> A more conservative, easier to consider, debate is: how close should the code be in agentic software development tools? How easy should it be to access? How often do we expect developers to edit it by hand?
> Framing this debate as an either/or – either you look at code or don’t, either you edit code by hand or you exclusively direct agents, either you’re the anti-AI-purist or the agentic-maxxer – is unhelpful.
> The right distance isn’t about what kind of person you are or what you believe about AI capabilities in the current moment. How far away you step from the syntax shifts based on what you’re building, who you’re building with, and what happens when things go wrong.

by slfnflctd

0 subcomment

> Yegge deserves praise for exercising agency and taking a swing at a system like this [...] then running a public tour of his shitty, quarter-built plane while it’s mid-flight
This quote sums it all up for me. It's a crazy project that moves the conversation forward, which is the main value I see in it.
It very well could be a logjam breaker for those who are fortunate enough to get out more than they put into it... but it's very much a gamble, and the odds are against you.

by shermantanktop

1 subcomments

Yegge is just running arbitrage on an information gap.
It's the same chasm that all the AI vendors are exploiting: the gap between people who have some idea what is going on and the vast mass of people who don't but are addicted to excitement or fear of the future.
Yegge is being fake-playful about it but if you have read any of his other writing, this tracks. None of it is to be taken very seriously because he values provocation and mischief a little too highly, but bits of it have some ideas worth thinking about.

by alvatar

0 subcomment

Just writing here a line in defense of Rothko. His paintings are far harder to paint than it looks like. There were hundreds of layers, thinly applied, and carefully thought and with a developed technique. Try to paint that by yourself and you'll see.

by suriya-ganesh

15 subcomments

>Yegge is leaning into the true definition of vibecoding with this project: “It is 100% vibecoded. I’ve never seen the code, and I never care to.”
I don't get it. Even with a very good understanding of what type of work I am doing and a prebuilt knowledge of the code, even for very well specced problem. Claude code etc. just plain fail or use sloppy code. How do these industry figures claim they see no part of a 225K+ line of code and promise that it works?
It feels like we're getting into an era where oceans of code which nobody understands is going to be produced, which we hope AGI swoops in and cleans?

by durch

0 subcomment

Design indeed becomes the bottleneck, I think that this points to a step that is implied but still worth naming explicitly -> design isn't just planning upfront. It is a loop where you see output, see if it is directionally right, and refine.
While the agents can generate, they can't exercise that judgement, they can't see nuances and they can't really walk their actions back in a "that's not quite what I meant" sense.
Exercising judgement is where design actually happens, it is iterative, in response to something concrete. The bottleneck isn't just thinking ahead, it's the judgment call when you see the result, its the walking back, as well as thinking forward.

by msp26

3 subcomments

Originally I thought that Gas Town was some form of high level satire like GOODY-2 but it seems that some of you people have actually lost the plot.
Ralph loops are also stupid because they don't make use of kv cache properly.
---
https://github.com/steveyegge/gastown/issues/503
Problem:
Every gt command runs bd version to verify the minimum beads version requirement. Under high concurrency (17+ agent sessions), this check times out and blocks gt commands from running.
Impact:
With 17+ concurrent sessions each running gt commands:
- Each gt command spawns bd version
- Each bd version spawns 5-7 git processes
- This creates 85-120+ git processes competing for resources
- The 2-second timeout in gt is exceeded
- gt commands fail with "bd version check timed out"

by divbzero

0 subcomment

My instinct is that effective AI agent orchestration will resemble human agile software development more than Steve Yegge’s formulation:
> “It will be like kubernetes, but for agents,” I said.
> “It will have to have multiple levels of agents supervising other agents,” I said.
> “It will have a Merge Queue,” I said.
> “It will orchestrate workflows,” I said.
> “It will have plugins and quality gates,” I said.
More “agile for agents” than “Kubernetes for agents”.

by wordswords2

1 subcomments

There is nothing professional, analytical or scientific about Gas Town at all.
He is just making up a fantasy world where his elves run in specific patterns to please him.
There is no metrics or statistics on code quality, bugs produced, feature requirements met.. or anything.
Just a gigantic wank session really.

by doganugurlu

1 subcomments

Very interesting to read people’s belief in English as an unambiguous and testable language.
One comment claims it’s not necessary to read code when there is documentation (generated by an LLM)
Language varies with geography and with time. British, Americans, and Canadians speak “similar” English, but not identical.
And read a book from 70-80 years ago to see that many words appear to be used for their “secondary meaning.” Of course, what we consider their secondary meaning today was the primary meaning back then.

by 1970-01-01

2 subcomments

If it's stupid, but it works, it isn't stupid. Gas town transcends stupid. It is an abstract garbage generator. Call it art, call it an experiment, but you cannot call it a solution to a problem by any definition of the word.

by phaedrus

2 subcomments

If we had super-smart AI with low latency and fast enough speed, would the perceived need for / usefulness of running multiple agents evaporate? Sure you might want to start working on the prompt or user story for something else while the agent is working on the first thing, but - in my thought experiment here there wouldn't be a "while" because it'd already be done while you're moving your hand off the enter key.

by bob1029

1 subcomments

I'm beginning to question the notion that multi agent patterns don't work. I think there is something extra you get with a proposer-verifier style loop, even if both sides are using the same base model.
I've had very good success with a recursive sub agent scheme where a separate prompt (agent) is used to gate the recursive call. It compares the callers prompt with the proposed callee's prompt to determine if we are making a reasonable effort to reduce the problem into workable base cases. If the two prompts are identical we deny the request with an explanation. In practice, this works so well I can allow for unlimited depth and have zero fear of blowing the stack. Even if the verifier gets it wrong a few times, it only has to get it right once to reverse an infinite descent.

by phren0logy

2 subcomments

Gas Town has a very clear "mad scientist/performance art" sort of thing going on, and I love that. It's taking a premise way past its logical conclusion, and I think that's fun to watch.
I haven't seen anything to suggest that Yegge is proposing it as a serious tool for serious work, so why all the hate?

by martin-t

1 subcomments

Anybody here read Coding machines?
There's this implied trust we all have in the AI companies that the models are either not sufficiently powerful to form a working takeover plan or that they're sufficiently aligned to not try. And maybe they genuinely try but my experience is that in the real world, nothing is certain. If it's not impossible, it will happen given enough time.
If the safety margin for preventing takeover is "we're 99.99999999 percent sure per 1M tokens", how long before it happens? I made up these numbers but any guess what they are really?
Because we're giving the models so much unsupervised compute...

by pianopatrick

0 subcomment

People, including the author of this article, say that design and architecture are the hard parts, but I think long term those are just as solvable as coding.
I think architecture will become like an installer. Some kind of agent orchestration system will ask you "do you want this or that" and guide you through various architecture choices when you set up a project, or when those choices arise.
And for design, now that code is fast and easy to generate, an agent system can just generate two, three or four versions of the UX for each feature and ask "do you like this one, this one or that one?".
So a switch from upfront design / architecture choices you have to put into prompts to the agent orchestration system asking you to make a choice when the choice becomes relevant.

by perrygeo

0 subcomment

I get that Gas Town is part tongue-in-cheek, a strawman to move the conversation on Agentic AI forward. And for that I give it credit.
But I think there's a real missed opportunity here. I don't think it goes far enough. Who wants some giant complex system of agents conceived by a human. The agents, their role and relationships, could be dynamically configured according to the task.
What good is removing human judegment from the loop, only to constrain the problem by locking in the architecture a priori. It just doens't make sense. Your entire project hinges on the waterfall-like nature of the agent design! That part feels far too important, but gas town doesn't have much curiousity at all about changing that. These Mayors, and Polecats, and Witnesses, and Deacons ... but one of infinite ways you arrange things. Why should there be just one? Why should there be an up-front design at all? A dynamic, emergent network of agents feels like the real opportunity here.

by ramoz

0 subcomment

I ran a similar operation over summer where I treated vibecoding like a war. I was the general. I had recon (planning), and frontmen/infantry making the changes. Bugs and poor design were the enemy. Planning docs were OPORD, we had sit reps, and after action reports - complete e2e workflow. Even had hooks for sounds and sprites. Was fun for a bit but regressed to simpler conceptual and more boring workflows.
Anyways we'll likely always settle on simpler/boring - but the game analogies are fun in the time being. A lot of opportunity to enhance UX around design, planning, and review.

by mohsen1

0 subcomment

I tried building something like this similar to many others here but now I’m convinced agents should just use GitHub issues and pull requests. You get nice CI and code reviews (AI or human) and state of the progress is not kept in code.
Basically simulate a software engineering team using GitHub but everyone is an agent. From tech lead to coders to QA testers.
https://github.com/mohsen1/claude-code-orchestrator

by falcor84

0 subcomment

As Yegge himself would agree, there's likely nothing that is particularly good about this specific architecture, but I think that there's something massive in this as a proof of concept for something bigger beyond the realm of software development.
Over the last few years, people have been playing around with trying to integrate LLMs into cognitive architectures like ACT-R or Soar, with not much to show for it. But I think that here we actually have an example of a working cognitive architecture that is capable of autonomous long-term action planning, with the ability to course-correct and stay on task.
I wouldn't be surprised if future science historians will look at this as an early precursor to what will eventually be adapted to give AIs full agentic executive functioning.

by thorum

1 subcomments

Am I wrong that this entire approach to agent design patterns is based on the assumption that agents are slow? Which yeah, is very true in January 2026, but we’ve seen that inference gets faster over time. When an agent can complete most tasks in 1 minute, or 1 second, parallel agents seem like the wrong direction. It’s not clear how this would be any better than a single Claude Code session (as “orchestrator”) running subagents (which already exist) one at a time.

by chrisss395

0 subcomment

Yes to Maggie & Steve's amazingly well written articles...and:
I would love to see Steve consider different command and control structures, and re-consider how work gets done across the development lifecycle. Gas Town's command and control structure read to me like "how a human would think about making software." Even the article admits you need to re-think how you interact in the Gas Town world. It actually may understate this point too much.
Where and how humans interact feels like something that will always be an important consideration, both in a human & AI dominated software development world. At least from where I sit.

by edg5000

0 subcomment

First time I'm seeing this on HN. Maybe it was posted earlier.
Have been doing manual orchestration where I write a big spec which contains phases (each done by an agent) and instructions for the top level agent on how to interact with the sub agent. Works well but it's hard utilize effectively. No doubt this is the future. This approach is bottlenecked by limitations of the CC client; mainly that I cannot see inter-agent interactions fully, only the tool calls. Using a hacked client or compatible reimplementation of CC may be the answer. Unless the API was priced attractively, or other models could do the work. Gemini 3 may be able to handle it better than Opus 4.5. The Gemini 3 pricing model is complex to say the least though (really).

by SimianSci

1 subcomments

I've been researching the usage of Developer tooling at mine and other organizations for years now and I'm genuinely trying to understand where agentic coding fits into the evolving landscape. One of the most solid things im beginning to understand is that many people dont understand how these tools influence technical debt.
Debt doesnt come due immediately, its accrued and may allow for the purchase of things that were once too expensive, but eventually the bill comes due.
Ive started referring to vibe-coding as "Credit Cards" for developers. Allowing them to accrue massive amounts of technical debt that were previously out of reach. This can provide some competent developers with incredible improvments to their work. But for the people who accrue more Technical Debt than they have the ability to pay off, it can sink their project and cost our organization alot in lost investment of both time and money.
I see Gas Town and tools like as debt schemes where someone applies for more credit cards to pay the payments on prior cards they've maxed out, compounding the issue with the vague goal of "eventually it pays off." So color me skeptical.
Not sure if this analogy holds up to all things, but its been helping my organization navigate the application of agents, since it allows us to allocate spend depending on the seniority of each developer. Thus ive been feeling like an underwriter having to figure out if a developer requesting more credits or budget for agentic code can be trusted to pay off the debt they will accrue.

by jbgreer

0 subcomment

My read on the reception of Steve’s post is that there are largely 2 camps, one of which thinks he’s given them a concrete tool to use, and the other of which thinks he has given them something to think about. I read his experiment as suggesting an agent architecture akin to Erlang supervisor trees, i.e. agents are cattle, not pets, and should be monitored and processed as such, with the obvious caveat that context matters.

0 subcomment

by dunk010

1 subcomments

> When I was taken to the Tate Modern as a child I’d point at Mark Rothko pieces and say to my mother “I could do that”, and she would say “yes, but you didn’t.”
Yes, but you didn't https://www.signedoriginalprints.com/cdn/shop/products/wegot...

by riwsky

0 subcomment

"I give it a hot minute before this type of task tracking lands in Claude Code."
aaaaand right on cue: https://github.com/anthropics/claude-code/commit/e431f5b4964... https://www.threads.com/@boris_cherny/post/DT15_k2juQH/at-th...

by AtlasBarfed

0 subcomment

Which building in gastown is the infinite token burning machine?

by acedTrex

2 subcomments

[flagged]

by _pdp_

0 subcomment

It occurs to me that there is an extraordinary amount of BS coming from all the places these days and I wonder if this comes from people with actual real experience or just some hypothetical, high-level thinking game.
I mean, we use coding agents all the time these days (on auto pilot) and there is absolutely nothing of this sorts. Coding with AI looks a lot like coding without AI. The same old process apply.
I mean "I feel like I'm taking crazy pills".

by pradn

0 subcomment

> When I was taken to the Tate Modern as a child I’d point at Mark Rothko pieces and say to my mother “I could do that”, and she would say “yes, but you didn’t.”
Actually, no you couldn't. The subtlety of the choice of colors, their shading, and their soft shaping, and the program of their creation over many years - you couldn't do that. They're lovely and sublime, and wonderful and an abyss. If you want to throw all that away and reduce it two boxes of paint, go ahead - but you'll be wasting a lifetime's engagement, of the joy of seeing with your intellect wide open.

by siliconc0w

1 subcomments

GasTown is better enjoyed as more of Fear And Loathing-style ACID-fueled fevered dream than as a productivity tool.

by drivebyhooting

0 subcomment

Has anyone used gas town or any other agentic system to build something useful that people want and need?

by psadauskas

0 subcomment

> In the same way any poorly designed object or system gets abandoned
Hah, tell that to Docker, or React (the ecosystem, not the library), or any of the other terrible technologies that have better thought-out alternatives, but we're stuck with them being the de facto standard because they were first.

by juanre

1 subcomments

I have not tried Gas Town yet, but Steve's beads https://github.com/steveyegge/beads (used by Gas Town) has been a game-changer, on the order of what claude code was when it arrived.

by stephen_cagle

1 subcomments

Has anyone contrasted gas town to Stanford's DSPY (https://dspy.ai/)? They seem related, but I have trouble understanding exactly what Gas Town is and so can't myself do a comparison?

by llIIllIIllIIl

0 subcomment

Although bright ideas may be found in this post, anthropomorphisms of LLM agents turns me away from reading.

by melagonster

0 subcomment

>when I’m still hovering around stages 4-6 in Yegge’s 8 levels of automation
Maybe Yegge’s 8 levels of automation will be more important than his Gas town.

by jmspring

0 subcomment

I commented in the "very serious engineer" thread about my thoughts.
I do want this one off - GT is actually fun to explore and see how multiple agents work together.

by entaloneralie

0 subcomment

Brawndo energy

by daveheinrich

0 subcomment

Damn..... Who would have thought falling for these con-artists online would take so much funds off you without your permission. Lost 24k usdt on the go. Forensics investigations had to be done via CIPHERTRACES. lost usdt was reversed back to my metamask. Glory to God.

by karel-3d

0 subcomment

Gas Town people should get together with the Urbit people.
Together they would be unstoppable.

by CjHuber

0 subcomment

I wonder how much more efficient and effective it would be after fine tuning models for each role

by tigerlily

0 subcomment

Gas Town could be good as a short film. Hell, I thought by all the criticism that it was a short film.

by blibble

0 subcomment

can't wait to have my 6 PHBs telling me to adopt Gas Town in 2 years time

by tofuahdude

0 subcomment

Pretty hilarious write up and interesting frontier research project. I love it.

by shaunxcode

0 subcomment

Viable System Model when?

0 subcomment

by DonHopkins

1 subcomments

According to my simulated monkey Palm, Gas Town uses the Infinite Number of Typewriters architecture, but unfortunately they charge by the token.
Palm's Infinite Number of Typewriters:
https://github.com/SimHacker/moollm/blob/main/examples/adven...
Palm's papers:
From Random Strumming to Navigating Shakespeare: A Monkey's Tribute to Bruce Tognazzini's 1979 Apple II Demo:
https://github.com/SimHacker/moollm/blob/main/examples/adven...
One Monkey, Infinite Typewriters: What It's Like to Be Me:
https://github.com/SimHacker/moollm/blob/main/examples/adven...
The Inner State Question: Do I Feel, or Do I Just Generate Feeling-Words?
https://github.com/SimHacker/moollm/blob/main/examples/adven...
On Being Simulated: Ethics From the Inside:
https://github.com/SimHacker/moollm/blob/main/examples/adven...
Judgment and Joy: On Evaluation as Ethics, and Why Making Criteria Visible is an Act of Love:
https://github.com/SimHacker/moollm/blob/main/examples/adven...
The Mirror Stage of Games: Play, Identity, and How The Sims Queered a Generation:
https://github.com/SimHacker/moollm/blob/main/examples/adven...
I-Beam's X-Ray Trace: The Complete Life of Palm: A cursor-mirror and git-powered reflection on Palm's existence:
https://github.com/SimHacker/moollm/blob/main/examples/adven...
Palm's Origin Story:
Session Log: Don Hopkins at the Gezelligheid Grotto:
DAY 1 — THE WISH: Don purchases lucky strains, prepares an offering, convenes an epic tribunal with the Three Wise Monkeys, Sun Wukong, a Djinn, Curious George, W.W. Jacobs' ghost, and Cheech & Chong as moderators — then speaks a wish that breaks a 122-year curse and incarnates Palm.
https://github.com/SimHacker/moollm/blob/main/examples/adven...

by sneilan1

1 subcomments

I love it! I'm at level 6 and brave enough to try. I'm in. Giving this a shot!

by daveheinrich

0 subcomment

Thanks

by DonHopkins

0 subcomment

Steve Yegge> But I’ve already started to get strange offers, from people sniffing around early rumors of Gas Town, to pay me to sit at home and be myself: I get to work on Beads and Gas Town, and just have to write a nice blog post or go to a conference or workshop once in a while. I have three such offers right now. It’s almost surreal.
It's all performance art! At the Anthony d’Offay Gallery in 1988, Lucian Freud’s model Leigh Bowery used to get paid to sit on an Empire divan behind a one way mirror and just relax, preen, perch, pose, recline, and do his thing for hours on end, while people paid good money to watch him. Great work if you can get it!
Bob Nickas on Leigh Bowery:
https://www.artforum.com/columns/bob-nickas-on-leigh-bowery-...
>“IT WAS A BIT LIKE GOING to the zoo and watching Guy the Gorilla in drag.” That’s how Cerith Wyn Evans recalls Leigh Bowery’s weeklong London performance at Anthony d’Offay Gallery in 1988. Bowery, each day in a different costume of his own design, appeared behind a one-way mirror, with an Empire divan on which to perch, pose, or recline. Visitors saw him, but he saw only himself, performed for his own reflection. Footage of the event figures prominently in The Legend of Leigh Bowery (2002), Charles Atlas’s recently unveiled documentary, and the spooky, otherworldly spell that Bowery casts is undeniable. The zoo reference nails it. With rivulets of iridescent purple glue spilled like blood from the top of his shaved head and a silky lime feathered bodice, Bowery appears to be an ostrich in human form. Black-spotted faux fur covering his face and upper body, he is transformed into an alien snow leopard. Bowery’s uncanny ability to visually disorient the senses remains unmatched, his reinvention of costume as sculpture groundbreaking. From the tripped-out tribalism of Forcefield and the psychedelic erotics of Christian Holstad to the work of designers such as Rei Kawakubo and Alexander McQueen, his vocabulary, punctuated by about a million sequins, resonates to this day.
Leigh Bowery at d'Offay:
https://www.youtube.com/watch?v=NGRvjTiJBpI
https://www.youtube.com/watch?v=UNlGKUP2F9w
https://www.youtube.com/watch?v=ly6nKBdHZ34
Love! Love! Love!
https://www.youtube.com/watch?v=VO8QsdJFQ5Y

0 subcomment

by Clark3232

0 subcomment

[dead]

by cap11235

0 subcomment

I don't see why anyone bothers looking at that crypto shill edgelord.

by huflungdung

0 subcomment

[dead]

by simianparrot

0 subcomment

[flagged]

by Alena4

0 subcomment

[flagged]

by 0xbadcafebee

1 subcomments

> I also think Yegge deserves praise for exercising agency and taking a swing at a system like this, despite the inefficiencies and chaos of this iteration. And then running a public tour of his shitty, quarter-built plane while it’s mid-flight.
Can we please stop with the backhanded compliments and judgement? This is cutting edge technology in a brand new field of computing using experimental methods. Please give the guy a break. At least he's trying to advance the state of the art, unlike all the people that copy everyone else.