The Shape of Things to Come

I smell an architecture astronaut getting high on abstractions. Taken far enough, it starts to look like psychedelic ramblings of a 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.

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 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.

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):

https://github.com/smol-env/smol

works just fine

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 ;-(

> 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.

But how much does the game bring in?

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).

I need to give this to an AI and ask it, "Does this post contain any information which will help me use Pi Agent Harness better?"

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.

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.

This guy's under a heavy AI psychosis

> 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.