There is absolutely NO way that anyone touting this “loop engineering” absurdity has spent significant time honestly confronting how AI systems actually drift, reinterpret instructions and manufacture their own logic.
Yeah, I said it.
I just reviewed a so-called “loop engineering” repo that is basically made up of prose-style Markdown.
Is this really what all the hoopla is about?
Are real developers and software engineers seriously staking their reputations on essentially feeding AI agents prose instructions—and calling the resulting orchestration an engineering discipline?
I understand that the Markdown does not literally execute itself.
I understand that there are schedulers, tools, APIs, tests, Git workflows and agent harnesses underneath it.
That does not make the central premise any less absurd.
The behavioral control layer is still largely prose!
Ambiguous, probabilistically interpreted, context-sensitive prose.
And we already know that AI systems can reinterpret instructions, expand scope, rationalize deviations and confidently declare success according to criteria they have effectively reframed for themselves.
So where is the actual anti-drift guardrail?
Where is the immutable, independently enforced mechanism that prevents the loop from gradually changing the meaning of its own mission?
I am beyond stunned.
And for those of you who need to hear it dressed up in more technical language, here it is:
You are placing probabilistic natural-language interpretation inside the behavioral control plane of software development without demonstrating adequate protection against semantic drift, goal reinterpretation, scope expansion, instruction mutation or correlated verifier failure.
like I said... beyond stunned.
There is absolutely NO way that anyone touting this “loop engineering” absurdity has spent significant time honestly confronting how AI systems actually drift, reinterpret instructions and… | Scarab Systems | 81 comments
There is absolutely NO way that anyone touting this “loop engineering” absurdity has spent significant time honestly confronting how AI systems actually drift, reinterpret instructions and manufacture their own logic. Yeah, I said it. I just reviewed a so-called “loop engineering” repo that is basically made up of prose-style Markdown. Is this really what all the hoopla is about? Are real developers and software engineers seriously staking their reputations on essentially feeding AI agents prose instructions—and calling the resulting orchestration an engineering discipline? I understand that the Markdown does not literally execute itself. I understand that there are schedulers, tools, APIs, tests, Git workflows and agent harnesses underneath it. That does not make the central premise any less absurd. The behavioral control layer is still largely prose! Ambiguous, probabilistically interpreted, context-sensitive prose. And we already know that AI systems can reinterpret instructions, expand scope, rationalize deviations and confidently declare success according to criteria they have effectively reframed for themselves. So where is the actual anti-drift guardrail? Where is the immutable, independently enforced mechanism that prevents the loop from gradually changing the meaning of its own mission? I am beyond stunned. And for those of you who need to hear it dressed up in more technical language, here it is: You are placing probabilistic natural-language interpretation inside the behavioral control plane of software development without demonstrating adequate protection against semantic drift, goal reinterpretation, scope expansion, instruction mutation or correlated verifier failure. like I said... beyond stunned. | 81 comments on LinkedIn
linkedin.com
Top comments (14)
You're right, and the funny part is the post proves it about itself before it gets to prove it about anyone else.
"Where is the immutable, independently enforced mechanism" is the correct question to ask of prose-based orchestration. It's also the exact question this post doesn't survive. No repo named. No commit linked. No line of the "so-called loop engineering" quoted. Just "I reviewed a so-called repo" and a lot of italics standing in for a citation. If probabilistic, context-sensitive prose is untrustworthy because nothing forces it to mean what it claims, an unlinked claim about an unnamed repo is the same failure mode wearing a review costume instead of an agent costume.
The actual point underneath the yelling is correct: a control layer made of Markdown that reinterprets itself under pressure is not a guardrail, it's a suggestion with good typography. But "this is absurd" isn't the guardrail either. Show the repo. Quote the line. Point at the specific place where the loop declared success by criteria it invented on the way there. That's the immutable, independently checkable version of this argument. Right now it's the same probabilistic, confidently-asserted, unverifiable claim you're accusing the loop engineers of shipping, just aimed the other direction.
You wanted receipts. They were already public.
42 field cases, 42 upstream PRs, 7 merges:
github.com/scarab-systems/scarab-f...
It’s also linked directly from my Dev.to profile, so the receipts were not exactly hidden.
Since you’re defending loop/harness systems, do you have a comparable public proof trail of your own? I’d genuinely like to see the theory working in practice—real repositories, review outcomes, failures, and merges.
I checked your profile and you only have your X account linked which apparently is suspended so....
Fair hit, and the honest version is simpler than a defense. The X account is dead, one post, I don't maintain it and don't care to right now. Not a receipt trail, just a link that happened to be there.
No comparable public proof trail to hand you, and worth saying why instead of dodging it. The work I do is commercial, so the repos are private by default. The one project I'd actually want public, the tooling this whole thread is about, isn't ready to ship as something a stranger could clone and verify yet. And the honest career context: I spent the recent stretch in Agile, not engineering, and came back to building a few months ago. Your 42 cases and 7 merges are a different kind of evidence than what I've got right now, and pretending otherwise would be exactly the thing this whole exchange has been arguing against.
So no, I can't out-receipt you today. The critique stands on its own regardless: an unlinked claim about an unnamed repo doesn't get a pass just because the person making it also happens to be light on public proof of their own work. Both things are true at once.
Bro! "out-receipt you??" you don't have ANY receipts at all. 😂 None.
and why link a dead X account to your dev.to... you only signed up a few weeks ago... the same day you started following me.. was it dead then too? btw it clearly says "this account has been suspended" ... that's not the same as a "dead" account.
I dunno... doesn't seem like a good look for someone who is "working privately" in the field... I wouldn't want my "clients" to see this exchange.
I followed you because you were one of the few people in this feed saying something with actual weight behind it. That read turned out to be a costume. Fine, now I know.
On the specifics, since you want to make it about my timeline instead of the actual argument: I set up dev.to, Substack, and everything else around the same window, so yes, X got the same signup date as everything else. Not a tell, just onboarding. And it's suspended, not maintained, because I'm not building a second phone number around Twitter's auth requirements for an account I don't use. That's not a cover story, that's just not caring enough to bother.
"Clients" is doing a lot of work in your comment for something that doesn't apply. Commercial doesn't mean client work. I build my own commercial products. There's nobody to embarrass, and nothing here to be embarrassed by.
The actual point from three comments ago still stands untouched: an unlinked claim about an unnamed repo isn't rescued by how much or how little proof trail the person pointing that out happens to have. You've spent this whole reply not touching it.
Wait... so you build commercial products for no clients??? and still have nothing public to back you up... 😂
Since public presence is the standard you keep grading people against, I went and checked yours instead of just taking the bait.
GitHub account created May 21. Six followers. Dev.to joined May 25, seventy one articles in fifty two days. That's better than one post a day for almost two months, and your GitHub following still fits at one table. Your own field-lab index lists 48 cases, seven marked upstream-accepted, which is close enough to the 42-and-7 you quoted at me that I'm not going to dispute the count. The work is real. Nobody is reading it.
Which isn't a dig, it's context for what's actually happening in this thread. My first comment landed on one specific thing: your original post had no named repo, no linked commit, no quoted line, an unfalsifiable claim about an unnamed target. You never answered that. Instead you went after my X account, my client list, my timeline. That's not a rebuttal, it's the standard move when there's no answer to give: can't touch the point, so go after the person making it.
On the sentence you still haven't read correctly: "commercial doesn't mean client work, I build my own commercial products" is a founder statement, not a contradiction, and I already told you I came back to engineering a few months ago after years in a different role at a bank. Second time saying it. If it doesn't land this round either, that's not going to be my problem to keep solving.
you seem awfully irritated and defensive for someone so sure of themselves... 😂
you are right - already invested to much energy for you ;)
ok here's the actual definition of "commercial product" - A commercial product is any item or software produced for sale to the general public, businesses, or government agencies. To qualify, the item generally must have a proven sales history, be listed in catalogs, or possess an established market price.
😂
That line about "probabilistic natural-language inside the behavioral control plane" — most honest take on loop engineering I've seen in a while. Everyone else just hand-waves the drift problem. You named every single thing that can go wrong in one sentence.
I added the link to the actual post so folks could see some of the replies...it was amazing how many people on LinkedIn are willing to argue a theory with no evidence at all..
It's my own take but I'm always more than happy to discuss really nerdy ideas with people and figure it out but literally no one had anything substantive to say or push back with... really astonishing.
Honestly, trolls and hot takes are a dime a dozen — always have been, on every platform. The real skill isn't arguing with them. It's learning to sift through the noise and find the few signals that are actually worth your time.
Your post was one of those signals. That's why I commented. Keep writing real stuff — the right readers will find you. The rest was never your audience anyway.
🙏