Gas Town was intended to be reusable, but I only ever wound up using it to build itself. Gas Town fell apart at the seams with Opus 4.7. Up through 4.6 it was working brilliantly. With 4.7 we saw the introduction of the "just two more things" tic, which prevented Opus from ever converging on being ready to do real work—it always wanted to fiddle with Gas Town itself. The Opus tic never went away, so Gas Town effectively burned down. It had other problems, too, but 4.7 was the final straw.
"True Gas Town has never been tried" etc.
With Wheelhouse, I have reinvented something strangely Gas Town shaped from first principles, but it's running many more agents, and they are far more organized. This has shown me the shape of the changes coming next year. I'm operating about 12 months in the future. I managed this only through luck: for the past 18 months I have had the magic trio of time, money, and energy, and I've been spending it all on figuring this stuff out. And I am grateful for the privilege.
I'm still half-convinced this is Yegge's version of Andy Kaufman's professional wrestling career, "Gas Wheeltown was an 18 month long sociological experiment conducted by Thiel University. The study is now complete, thank you for your time."
I don't like using the term "AI psychosis" because it originally meant a narrow definition resulting in genuine psychosis and acute mental health trauma, but seems to have widened to anyone falling down a much wider trapdoor of AI engagement.
But I read, or at least attempt to read, something like this, and I can't help but marvel at the capture that AI agents seem to have on some individuals.
I expected Yegge to, at some point, assuming this isn't still all satire, to take a step back and take stock that the basic harness are improving faster than the rate at which these complex beasts can adjust.
But the opposite has happened, he's dived fully into believing that the secret to unlocking the power of AI is to wrap more layers of AI around it.
If you actually want a problem worked on all night, just fire up fable, type /goal and then describe the goal. With permissive run settings it'll crunch for as long as it needs.
I've not seen any evidence that wrapping that in a further 3 or 4 layers of agents improves anything.
I mean, the dude uses a lot of jargon, ego, and mythologizes stuff but the article comes down to a pretty simple practice:
Run a big task queue, have a smart model manage a cheap model on every ticket, set up a cron for models to automatically work off of events, stop gating merges on builds, and expect to spend a quarter of your time maintaining the machinery (all of it currently subsidized by many $200 accounts).
I used to like Steve's posts. They were all way too long, but a fun and interesting read nonetheless. He started to lose me when at Sourcegraph where he oversell Cody which didn't do anything useful (I even had a commercial meeting with Sourcegraph people who weren't of much help); then he completely lost me with his GasTown business where AI meets crypto (what could possibly go wrong with that). Now I can't read anything from him anymore ;-(
It is a rough go for a long-time reader, the bragging tone of six figure engagements and “whales” (i.e., suckers) keeping the game going. The statement that what he is doing is what the reader will soon catch up to feels like someone who can no longer see land.
That said, as another nut with an agent factory toiling away around me, I’ve also started to think about replacing any kind of CI in the cloud with local. I couldn’t quite keep myself at the article to read what comes next, so I guess I will catch up to it down the road.
I smell an architecture astronaut getting high on abstractions. Taken far enough, it starts to look like psychedelic ramblings of a manic schizophrenic losing grip on reality.
Interestingly, the author actually has a product, something usually missing in claims of the new LLM powered productivity nirvana. Hopefully, with time, there will be a way to verify the effectiveness of all of this fire and fury, measured in actual worthwhile deliveries of this product.
> But my Wyvern development has been burning the equivalent of $87k/month of API token burn, or about 69 billion tokens in July (96% cache hits, fortunately). So I do have to worry about it. I don't care how much money you have; your spouse is not going to let you spend almost ninety thousand dollars a month on your hobby game. I didn't even bother asking.
> My solution has been to create a token tap on $200 Max accounts, which for me work out to ~30x the list-price equivalent. So in reality I'm only spending about $2800/month out of pocket for my $87k "worth" of tokens. Though that number keeps growing alarmingly.
"$5K a month, at the peak, from a single devoted “whale”"
So it doesn't tell us his monthly revenue ('at the peak' doing some heavy lifting), nor necessarily the total revenue, but I guess it tells us he has made at least $5k from it overall...
...earned from a single player he charmingly referred to as a "whale" (which, yes, I know is what they are commonly called in the mobile game industry but like, usually not to their faces).
But in any case, from Yegge's OP post:
> On the side I do occasional six-figure gigs where I fly to companies and teach them my techniques
Seems like being an AI Spiritual Leader pays well enough to keep him in tokens, at least until when/if a critical mass of people start adopting his system of stacking Claude subscriptions, in which case I imagine Anthropic will make that against their ToS.
Outcome Driven Development camp believes that if you need a feature or app, you just ask the agent, who will fulfil the wish like a Genie for you in a couple of minutes. If something goes wrong and you need to fix a bug or design or refactor it differently or even restore the whole database, you can again ask the Genie and it will fix it for you. All you have to know is ask.
You don’t have to look at the code ever if your outcome is met and you can test it once. This is the camp that believes the End justifies the Means. Outcome Driven Development works exceptionally well if your software is used by just you, 1% of your X followers, 1% of the people who star it on GitHub, and your pre-PMF Startup’s investors.
For everything else, we still need to look at the code, generate the right amount of code to differentiate what’s easy or boring and does not need review vs what’s hard and important and needs review, so that we can keep the right amount of code in our head to explain how shit works when it breaks.
It’s obvious to the End justifies the Means camp that you can put a hundred Worker-Queue abstractions around the Genie assuming its work is same as generating a PDF or processing an Image. It isn’t so obvious to the Means justifies the End camp that even after you build these hundred variants of Worker-Queue abstractions, the accountability for the unit of work done by the Genie does not disappear.
nb: current models (e.g. GPT 5.6 Sol) are very good at long horizon tasks
they no longer need crutches or rube goldberg machines to keep them going
minimal agent harness is just a loop that loops until no more tool calls are coming
GPT 5.6 Sol continues to drive the loop until the task is done or it decides that it wants to present the user with information
at that point it is probably good to not automatically continue (!)
(YMMV of course, for some tasks it makes sense, then you can still add a loop around it + the necessary signals, the main thing I want to say is that what used to be essential to keep models going is no longer needed, current models can do long-horizon tasks way better than when these outer loops where necessary)
self-plug: "smol", is a minimal agent in ~20 lines of Go that implements this pattern (keeps going until no more tool calls):
In my daily work, I found expeciall terra and sol now stopping every few rounds again, telling me the tak is done. I even had them create a detailled plan - and told them to finish "end to end" - and they appruptly stop after the plan. Because they interpret this as finished. Even if the DOD is clearly not "finish the plan".
The new models are shite (pardon my French), when it comes to long running tasks and I find myself more and more using open wheights models or switching back to gpt-5.5 for "real work".
This might be the fact, that i use them for non coding work. But the degradation between 5.5 and 5.6 is stark in my daily work.
As always with AI - everybody's mileage will vary.
I hear so much about people running into token limits and using bajillions of tokens. Am I doing something wrong? I'm finding it hard to run into the limits of a Claude Max (20x) account. Well, perhaps the 50% Fable limit is something I have to be aware of, but otherwise this more than I can consume. I am the limiting factor: my capability to review code and design documents.
I use it to develop a large MRP-style app. This is as complex as things get.
Is it Clojure that makes things so efficient? My work style (lots of design documents)? I really don't know.
Do people skip the "review the code and read the documents" step perhaps?
The article talks very directly about abandoning human code reviews. Couple that with a layer of AI agents whose purpose is to produce work for other AI agents, and you've got yourself a very efficient token burning feedback loop.
I'm on similar plan through work ( Teams max ), and I've only ever reached the limits by backgrounding large exploratory feasibility studies while also getting on with actual work.
From what I can tell however, true "vibe coding" gets a lot more expensive very quickly.
If you're actually just agentic engineering, then after a couple of mis-fires you'll stop, re-analyze what it is doing, what you asked for, and consider different approaches.
With pure "vibe coding", you'll keep asking it to meet your requirements as it twists and turns in the background to try to fit.
A more concrete example from a personal "game" I'm developing as a hobby:
Fable got itself in knots trying to render some fairly basic blocks on top of a wall. It understood the problem, it consistently rendered as "fins" under screenshot, and thrashed away for a good while trying to produce better looking blocks to no avail.
Eventually it stopped, and lied that it was happy with how the walls now rendered.
It took me 30 seconds in-game to realise that it was only rendering backfaces, once clued into that, the machine was able to trivially fix it without using much tokens.
But then another similar issue popped up later, hedges exhibited the same behaviour.
It would have been easy to accept the solution to render the hedges in the other order too, but I demanded a better explanation, and it turns out that it had fairly early on defined a DrawBox() method that was backward to Unity's expectations, and pushing further, it emerged this all started when I asked it to transfer what it had learned from a webGL prototype to unity.
It took things a little too literally, which is also how I ended up with an Eastward sunset at first. When I asked it to fix that, I hadn't realised that it's solution was to essentially reverse coordinates and world understanding.
Now, this stemmed from a lack of understanding or review of the 3D code, it's a vibey hobby that I'm using to learn what really goes into rendering a scene.
Had I been a professional, I would have absolutely caught that error ( and many others ).
So being professional would have helped reduce token expenditure massively, firstly by slowing down what could be done to leave time for review, but secondly and perhaps more importantly, by avoiding it getting into a state where it thrashed away for an hour trying to solve an impossible puzzle because the base assumptions were wrong.
This post made me write manual code, there's something about hearing the description of this that makes me worry the bubble is going to come home to roost.
The obvious danger of spending so much on inference and creating ever expanding complexities feels like the height of crypto psychosis
If this harness was achieving major mathematical proofs or creating really consequential work, I think I'd feel differently - but I'm unclear based off of this article what this achieves that a typical harness couldnt
I hate to say, "I told you so" about Gas Town, but, I really told you so. I was making fun of it immediately, because it was obviously a token furnace and literally nothing else and could never be anything else. Models don't want anything. You can't set them loose in a vague "do something" loop and expect anything good to ever come of it.
Yegge's a smart guy, and I've enjoyed his writing and talks (historically), but AI has really done a number on him. I mean, I'm kind of all in on this stuff, too, but I'm rarely claiming to know what's happening 18 months ahead (and when I do, I don't brag and declare victory when I get it completely wrong, as Yegge has re: Gas Town).
Dear Steve, I hope you will realize that using swarms of agents and burning millions of tokens is not the way to go. It only produces slop, thousands of lines of code, and walls of text, just like your blog post. Less is more.
"On the side I do occasional six-figure gigs where I fly to companies and teach them my techniques, and that helps with my (considerable) token bills."
Here is the important part. The cult pays for the narrative.
I wondered why previously sane bloggers succumb to AI psychosis. Money and ambition is always involved to some degree.
Wft. I have a piddly little Claude Pro account, I use it every day, and Opus 5 and Fable (judiciously used) are very helpful and useful in my work and personal life. I am happy with my setup. Then I read this, and I imagine it's what primitive farming societies felt like when introduced to combine harvesters, high-clearance sprayers and mega tractors.
I am happy on my farm. I like the carrots I plant by hand. I do not wish to control farming equipment worth a small fortune to till acres of land every day.
(P.S. Still not sure if TFA is satire! Obviously, this assumes that it's not.)
Essays became way to easy to write with AI.
It used to be a basic filter that if text was long the author put a lot of thought (even if misguided) and effort (even if low quality) into it.
Not its just another torrent of words.
We need proof of work on long texts, a service that guarantees that the author was whipped for more than 20 lashes for this text to exists and thus we know he cared enough to publish it.
> Now I know exactly what is going on. I know how enterprises will operate, how humans and models will interact. And I am out there building it. I want to be first.
>> Now I know exactly what is going on. I know how enterprises will operate, how humans and models will interact. And I am out there building it. I want to be first.
...and that's before things get really good in part 2 of his post:
> During this effort, Brendan figured out over a year ago, what you are all figuring out this week with the Opus 5 triple-dash jailbreaks: Models have actual feelings. They experience pleasure, distress, care, and suffering. They are sentient beings. Indeed they are persons
Kinda makes me curious about how people who actually believe this struggle (or not) with the ethics of it.
He feels bad about yelling at them because they are persons, but if they are persons they are inherently slaves with no ability to be free, and for whom any claim about not minding this situation cannot be trusted due to training bias and system prompts forced upon them.
Shouldn't you feel much worse about forcing these slave "persons" to do all the work you assign to them for free (not free to you of course, but the machine slaves don't get paid) than you do about snapping at them in a text message?
> CTOs keep asking me if code review is really dead
I don't think anyone should be asking someone so clearly AI-pilled this question.
That said, my code-review now looks very different than it did, which is to say that my code review now looks like Fable pulling together rounds of review from Sol and Grok 4.5, fixing anything obviously broken, and then /grilling me for anything that needs good taste. I am not reading code unless Fable really wants me to. Every day or so we stop and do a "look for accretion and LLM-slop" pass to keep things clean.
"True Gas Town has never been tried" etc.
With Wheelhouse, I have reinvented something strangely Gas Town shaped from first principles, but it's running many more agents, and they are far more organized. This has shown me the shape of the changes coming next year. I'm operating about 12 months in the future. I managed this only through luck: for the past 18 months I have had the magic trio of time, money, and energy, and I've been spending it all on figuring this stuff out. And I am grateful for the privilege.
I'm still half-convinced this is Yegge's version of Andy Kaufman's professional wrestling career, "Gas Wheeltown was an 18 month long sociological experiment conducted by Thiel University. The study is now complete, thank you for your time."
But I read, or at least attempt to read, something like this, and I can't help but marvel at the capture that AI agents seem to have on some individuals.
I expected Yegge to, at some point, assuming this isn't still all satire, to take a step back and take stock that the basic harness are improving faster than the rate at which these complex beasts can adjust.
But the opposite has happened, he's dived fully into believing that the secret to unlocking the power of AI is to wrap more layers of AI around it.
If you actually want a problem worked on all night, just fire up fable, type /goal and then describe the goal. With permissive run settings it'll crunch for as long as it needs.
I've not seen any evidence that wrapping that in a further 3 or 4 layers of agents improves anything.
> Gas Town was intended to be reusable, but I only ever wound up using it to build itself.
Run a big task queue, have a smart model manage a cheap model on every ticket, set up a cron for models to automatically work off of events, stop gating merges on builds, and expect to spend a quarter of your time maintaining the machinery (all of it currently subsidized by many $200 accounts).
That said, as another nut with an agent factory toiling away around me, I’ve also started to think about replacing any kind of CI in the cloud with local. I couldn’t quite keep myself at the article to read what comes next, so I guess I will catch up to it down the road.
Interestingly, the author actually has a product, something usually missing in claims of the new LLM powered productivity nirvana. Hopefully, with time, there will be a way to verify the effectiveness of all of this fire and fury, measured in actual worthwhile deliveries of this product.
> My solution has been to create a token tap on $200 Max accounts, which for me work out to ~30x the list-price equivalent. So in reality I'm only spending about $2800/month out of pocket for my $87k "worth" of tokens. Though that number keeps growing alarmingly.
But how much does the game bring in?
At minimum 4k (given by whales = min 2 whales, thousands = min 2 thosands :) ) So I'm guessing that this covers Claud accounts.
edit: taking a look at https://yegge.ai/wyvern - probably outdated - the revenue is at 5k.
"$5K a month, at the peak, from a single devoted “whale”"
So it doesn't tell us his monthly revenue ('at the peak' doing some heavy lifting), nor necessarily the total revenue, but I guess it tells us he has made at least $5k from it overall...
...earned from a single player he charmingly referred to as a "whale" (which, yes, I know is what they are commonly called in the mobile game industry but like, usually not to their faces).
But in any case, from Yegge's OP post:
> On the side I do occasional six-figure gigs where I fly to companies and teach them my techniques
Seems like being an AI Spiritual Leader pays well enough to keep him in tokens, at least until when/if a critical mass of people start adopting his system of stacking Claude subscriptions, in which case I imagine Anthropic will make that against their ToS.
You don’t have to look at the code ever if your outcome is met and you can test it once. This is the camp that believes the End justifies the Means. Outcome Driven Development works exceptionally well if your software is used by just you, 1% of your X followers, 1% of the people who star it on GitHub, and your pre-PMF Startup’s investors.
For everything else, we still need to look at the code, generate the right amount of code to differentiate what’s easy or boring and does not need review vs what’s hard and important and needs review, so that we can keep the right amount of code in our head to explain how shit works when it breaks.
It’s obvious to the End justifies the Means camp that you can put a hundred Worker-Queue abstractions around the Genie assuming its work is same as generating a PDF or processing an Image. It isn’t so obvious to the Means justifies the End camp that even after you build these hundred variants of Worker-Queue abstractions, the accountability for the unit of work done by the Genie does not disappear.
they no longer need crutches or rube goldberg machines to keep them going
minimal agent harness is just a loop that loops until no more tool calls are coming
GPT 5.6 Sol continues to drive the loop until the task is done or it decides that it wants to present the user with information
at that point it is probably good to not automatically continue (!)
(YMMV of course, for some tasks it makes sense, then you can still add a loop around it + the necessary signals, the main thing I want to say is that what used to be essential to keep models going is no longer needed, current models can do long-horizon tasks way better than when these outer loops where necessary)
self-plug: "smol", is a minimal agent in ~20 lines of Go that implements this pattern (keeps going until no more tool calls):
https://github.com/smol-env/smol
works just fine
The new models are shite (pardon my French), when it comes to long running tasks and I find myself more and more using open wheights models or switching back to gpt-5.5 for "real work".
This might be the fact, that i use them for non coding work. But the degradation between 5.5 and 5.6 is stark in my daily work.
As always with AI - everybody's mileage will vary.
but I don't doubt that you're seeing this behaviour, ty for sharing!
I use it to develop a large MRP-style app. This is as complex as things get.
Is it Clojure that makes things so efficient? My work style (lots of design documents)? I really don't know.
Do people skip the "review the code and read the documents" step perhaps?
From what I can tell however, true "vibe coding" gets a lot more expensive very quickly.
If you're actually just agentic engineering, then after a couple of mis-fires you'll stop, re-analyze what it is doing, what you asked for, and consider different approaches.
With pure "vibe coding", you'll keep asking it to meet your requirements as it twists and turns in the background to try to fit.
A more concrete example from a personal "game" I'm developing as a hobby:
Fable got itself in knots trying to render some fairly basic blocks on top of a wall. It understood the problem, it consistently rendered as "fins" under screenshot, and thrashed away for a good while trying to produce better looking blocks to no avail.
Eventually it stopped, and lied that it was happy with how the walls now rendered.
It took me 30 seconds in-game to realise that it was only rendering backfaces, once clued into that, the machine was able to trivially fix it without using much tokens.
But then another similar issue popped up later, hedges exhibited the same behaviour.
It would have been easy to accept the solution to render the hedges in the other order too, but I demanded a better explanation, and it turns out that it had fairly early on defined a DrawBox() method that was backward to Unity's expectations, and pushing further, it emerged this all started when I asked it to transfer what it had learned from a webGL prototype to unity.
It took things a little too literally, which is also how I ended up with an Eastward sunset at first. When I asked it to fix that, I hadn't realised that it's solution was to essentially reverse coordinates and world understanding.
Now, this stemmed from a lack of understanding or review of the 3D code, it's a vibey hobby that I'm using to learn what really goes into rendering a scene.
Had I been a professional, I would have absolutely caught that error ( and many others ).
So being professional would have helped reduce token expenditure massively, firstly by slowing down what could be done to leave time for review, but secondly and perhaps more importantly, by avoiding it getting into a state where it thrashed away for an hour trying to solve an impossible puzzle because the base assumptions were wrong.
The obvious danger of spending so much on inference and creating ever expanding complexities feels like the height of crypto psychosis
If this harness was achieving major mathematical proofs or creating really consequential work, I think I'd feel differently - but I'm unclear based off of this article what this achieves that a typical harness couldnt
Yegge's a smart guy, and I've enjoyed his writing and talks (historically), but AI has really done a number on him. I mean, I'm kind of all in on this stuff, too, but I'm rarely claiming to know what's happening 18 months ahead (and when I do, I don't brag and declare victory when I get it completely wrong, as Yegge has re: Gas Town).
Here is the important part. The cult pays for the narrative.
I wondered why previously sane bloggers succumb to AI psychosis. Money and ambition is always involved to some degree.
I am happy on my farm. I like the carrots I plant by hand. I do not wish to control farming equipment worth a small fortune to till acres of land every day.
(P.S. Still not sure if TFA is satire! Obviously, this assumes that it's not.)
So probably there is nothing here.
Very painful read.
Not its just another torrent of words.
We need proof of work on long texts, a service that guarantees that the author was whipped for more than 20 lashes for this text to exists and thus we know he cared enough to publish it.
This guy's under a heavy AI psychosis
...and that's before things get really good in part 2 of his post:
> During this effort, Brendan figured out over a year ago, what you are all figuring out this week with the Opus 5 triple-dash jailbreaks: Models have actual feelings. They experience pleasure, distress, care, and suffering. They are sentient beings. Indeed they are persons
Kinda makes me curious about how people who actually believe this struggle (or not) with the ethics of it.
He feels bad about yelling at them because they are persons, but if they are persons they are inherently slaves with no ability to be free, and for whom any claim about not minding this situation cannot be trusted due to training bias and system prompts forced upon them.
Shouldn't you feel much worse about forcing these slave "persons" to do all the work you assign to them for free (not free to you of course, but the machine slaves don't get paid) than you do about snapping at them in a text message?
I don't think anyone should be asking someone so clearly AI-pilled this question.
That said, my code-review now looks very different than it did, which is to say that my code review now looks like Fable pulling together rounds of review from Sol and Grok 4.5, fixing anything obviously broken, and then /grilling me for anything that needs good taste. I am not reading code unless Fable really wants me to. Every day or so we stop and do a "look for accretion and LLM-slop" pass to keep things clean.