The real problem with agentic coding

I understand the issue is very polarized. Despite grieving old programming, I think agentic coding is the future as I have already stated on another post on this website. However I think there is a real problem with agentic coding that I would like to ask your opinion about. I would like knowing how you are dealing with it. The issue is money. I'm not talking about company adoption. I'm talking about personal development and staying relevant. To do that, you need to try the new tools. If I had access, I would spend a lot of time on it. I don't enjoy vibecoding and wouldn't want to rely on a few monster .md files. I would like to experiment a lot with LLM wrestling and trying to get deterministic output with very well structured prompts. The problem is I cannot do that because I don't have $1,000 to spend on AI. I don't have enough money to get a beer without checking my wallet. I'm a student in Europe. This new revolution pisses me off because I am cut out from the game. Sure, I can still program by hand and I really enjoy it, but this is not the most efficient way to do it. It's like traveling on foot: if it's 10 minutes it's okay, if I have to walk for 2 hours and it would take 20 minutes with a car, it's a real problem. I'm wasting lot of time for no good reason. I don't have the hardware necessary to run good local models. I don't have the money to pay for AI subscriptions that are at their lowest price because of subsidies. Universities don't provide tokens to CS students at the moment. This is really the worst time to enter the industry, and it is not because of job positions: it is because of economic disparity.

>Using (and paying for) proprietary spyware crap made to create the surveillance state and transfer the power from the users to the corporation >Will never learn anything by letting the LLM "code" for him >Programming to "stay relevant" in the job market >Will make sloppy programs with security vulnerabilities Hello lainchan??

>>44240 > if I have to walk for 2 hours and it would take 20 minutes with a car, maybe i'm naive, but wouldn't those 2 hours become 1 hour with enough experience? and in that one hour, i imagine you'll end up with more extensible code that you're already familiar with since..you wrote it. i think experience evaluating and constructing stable code bases would be valued over operating the black box code-writing machine in any company the average lain would consider worth working for. why are you convinced otherwise?

What LLMs facilitate is not useful. Notice that no AI company has made a profit and that the only AI users making money are slop generators: people who spam app stores with low quality garbage. This fad will pass. OpenJDK and Clojure core forbid LLM contributions, so I've been learning Clojure.

>>44240 >I would like to experiment a lot with LLM wrestling and trying to get deterministic output with very well structured prompts. >trying to get deterministic output >with very well structured prompts. I think this is a wrong move. If you can run the LLM locally, you can tune the "heat" parameter to make it more or less deterministic- or even interfere more intricately and deliberately to constrain the range of paths taken through the model. I haven't kept up with LLM research much but I think there are some terms that refer to this sort of thing, maybe "insertions"? As for the topic, yeah they're trying to make us all obsolete. It isn't a fair race, and we can't trust in them to fail. The only option short of insurrection is to find leverage in the weird places they won't think to look.

I agree it is really the worst time to enter the industry but not quite for the same reasons. As a europoor even the 2010s would rarely pay a senior dev something that a junior in the US would even get out of bed for. But now with the IT sector being taken over by indians who think they're hot soykaf by prompting their AI agents to vibe code one hot mess after another and the short term thinking managers needing you to match that output, also Europe being on the decline on all fronts in general, I think you better come up with something smart for yourself. >>44243 is right, no agentic workflow is going to replace you knowing what the hell is even going on. I might have been lucky in a sense when I started while there wasn't even auto completion around. But considering the time critical nature of your situation I believe what might work best is if you use AI as a learning aid and boilerplate generator. A quantized 13b model on an RTX3060 or AMD equivalent could already suffice. Also look into high VRAM mini PCs with an AMD AI chip which comes at a cost but would still be more cost efficient than a graphics card of the same memory capacity. And for running LLMs, memory is the bottleneck. The thing that often eats a great portion of time is getting stuck on some stupid obscure detail which is much easier to solve with an LLM. The agentinc vibe coding the entire codebase thing looks impressive at first but that's where even the larger models struggle with to keep consistency and sanity. You're better off learning existing tools to generate repetitive code files via templates and keep tabs on the structure as you go. Feeding sections of it into an ai to generate the scaffold or common algos. It's stuff like having a basic outline or example being generated that speeds things up considerably but make sure to also have it explain everything and get proficient in the tools you're using. >>44244 >What LLMs facilitate is not useful Not with that lack of imagination it isn't. True, agentic coding is mostly a meme and I'm still skeptical on Google's demo of vibe coding an entire operating system in a day, but the tools are not to be underestimated as I've learnt multiple languages by now much faster than I would have if parsing the more in-depth knowledge together from sparse stackoverflow posts and not always so helpful documentation. As for the fad, I am afraid that the onslaught of software slop won't end any time soon. >>44245 So far I haven't found a way to control the model's internal query mechanism but you can start a new context with a summary of the old to continue where you've left off. And I won't share my ideas in the open to delay counter measures but they're not infallible, there's a lot of hubris and false promises to investors involved. But there is ways to outplay them. I trust in you to be smart enough to figure it out too. These tools are of a new era, but as OP, being restrained in resources and time requires resourcefulness and ingenuity.

>>44249 >I've learnt multiple languages by now much faster than I would have if parsing the more in-depth knowledge together from sparse stackoverflow post When I read this, I don't think "Wow, that's impressive. LLMs must be useful."; I think "It takes me a couple days to be productive in all but the most esoteric languages, like Forth and APL. What is meant by this?"

>>44266 >What is meant by this? All programming languages are equally powerful.

>>44240 All those who say they "vibe" code stuff never say how many times they have to correct the material. They'll probably insist that there's still a real gain even with the corrections included. But then why is that India is still managing to slot in so many Hindoo coders? You can't have both a reduction of jobs and yet a growing number of cheap coders who keep spewing broken code that makes you waste time and money. The real loss with agentic stuff: it's all feeling similar, or you never know where you data ends even if they tell you that assuming you pay a soykaf lot of cash it's not used to train IA. Finally, the only way for this to function is to hoard natural resources at the expense of people's quality of life and their habitat. Energy has already been artificially inflated, water is getting scarce, yet people come second to AI promises that don't deliver. Not a single one of these companies is making net gains but I don't think it matters because we'll be the collateral while banks and governments will bail them out as they're becoming way too entrenched into our industries to be removed. They're like fused parasites. They are all losing massive amounts of money and it's not improving because they have to invest in the construction of power plants that they want to own directly just to run even more of their IA monsters. >>44249 >Feeding sections of it into an ai to generate the scaffold or common algos. You should still verify the code manually or trust it blindly.

>I don't have the hardware necessary to run good local models Are you sure? I've seen people getting decent results with a 1080. You don't need something huge to get massive performance gains. AI acceleration is pretty clearly logarithmic in nature, going from X to 2X the model power gets you X + K the benefit, where K is some constant. Huge for small values of X, but not really work it for large values of X. It's why the entire AI industry is an absolute scam that will collapse the economy, their entire premise is that X MUST increase at all costs.

>>44421 Sorry: going from X to 2X the model power takes you from Y to Y + K benefit.

>>44241 Welcome to Lainchan's rotting, undead corpse, embrace the proprietary slop and scream in glee as we reach the next phase of technology's decline by accessibility.

>>44426 consoomers on a cyberpunk ib. what has happened...

>>44241 Ironically AI is finding like 10 vulnerabilities per week over the last few months in Linux kernel and other important software. Not saying it's writing great code or anything.

>>44428 Have you actually read these vulnerability reports? Like most CVE's, of them are just minor bugs that are dressed up like severe security vulnerabilities, and the actually significant ones are mostly done with significant human involvement. Meanwhile a tremendous amount of completely frivolous "vulnerability reports" has forced many open source projects to change their submission policies.

>>44429 I don't disagree with the core premise that most LLM identified bugs are slop and bug-fodder, but I have a fully automated RE->DBG pipeline that's made me roughly $15k my farming LPE bugs. I have similar functionality for web that's netted me some unauth SQL/SSRF and RCE in Apache products. Let's be real, though, Apache isn't really the hardest target out there. To the credit of your point, I still had to berate the LLM so stop over-exaggerating the severity of lame bugs like "debugger assisted privilege escalation" (lmao) and cookie based XSS on a POST-based request. I had to spend ~5h+ of work steering it, calling it dumb, and removing bullshit findings. I will say, though, the models I have tested are fuarrrking TERRIBLE at complex/multi-chain bugs and business logic bugs. There's miles and miles to go before I'll be worried about my job security.

>>44434 tf does this even mean?

>>44435 he's a spammer, report when you see that kind of post.

The real problem with agentic coding is that you'll get literally dumber if you do it. It's alright if you want to dedicate your energy elsewhere, but let's be honest, most people would just distract themselves with something shiny while AI does the lifting for them. I think it's understandable to do it to survive in the market, of course, if you produce slop to a slop industry and gets actual money from it, who gives a fuarrrk. It's like feeding soykaf to the pigs and getting paid for it. But an intermediate will downgrade to a novice in months and eventually you won't be able to call yourself a programmer anymore. Well, if you care about your pride as a professional and love what you do, you wouldn't do it. If you don't, like most people out there, then you probably don't belong here.

>>44241 >>44240 OP, there's something you should consider. INDUSTRIAL programming. Programming industrial-grade stuff is something they just cannot let AI to do. Besides, industrial programming requires you to know your stuff, because the price of error is high. Waaaaaaaaay too high to let aislop in. 1. Allen-Bradley still runs XX century soykaf 2. LADDER and alike are rather easy languages 3. SIEMENS' Simatic is somewhere in /y2k/ - it still ships 125 KB S7-1200 series CPUs for 400$+

>>44243 Doubled speed via experience? Hmph! Only if you started with flat feet (and aching/huring knees) but got the right special orthopedic shoes throughout the months "experience" (hence getting rid of the knee pain)

you don't even have enough ram for gemma 4 26b or Qwen3.6 35B at q4??

I guess people can use LLMs to declassify state secrets now?

>>44240 The philosophical failure with agentic coding is that it replaces more rigorous turing complete systems with fuzzy ones inherently prone to being incorrect. To solve agentic coding or LLMs would effectively be to solve thinking, which is either insane or dishonest to even attempt. >Sure, I can still program by hand and I really enjoy it You don't know what ur goal is or what's efficient because your conscious and subconscious self often disagree, and then you don't really understand what it was you were doing or trying to do until you do it, and not even then do you fully understand any of that. As above the concept of automated thinking is cope and in the age where anyone can access or easily generate information the new value is in being correct & having unique perspectives, the LLMs as they exist like the claude api are built to facilitate alienation from your cognitive labor, entrap you in that system, and gradually increase API fees or make you distrust yourself, because it is optimal for profit. Decide on your own course of action, read Marx, Foucault, study history, whatever. Not really a new problem, not an LLM problem, not an AI problem, not a coding problem.

>>44563 In medical tech there's a big push to do ai soykaf and boeing already let a plane crash by outsourcing to incompetent third worlders, they will do it as long as the average punishment does not exceed the profit or likely profit

>>44415 It is telling that vibe coding has not produced nearly as many tech products or services people find worth spending money on compared to methods employed in the 90s/2000s/2010s. If you aren't producing more soykaf the consumer wants it will collapse. A lot of this is downstream of public structures used by corporations which lead to incompetent, short sighted, stupid investors taking control. All the ai data centers will prob be retrofitted for sentiment analysis and surveillance when they finally admit no one wants it and they're hitting diminishing returns.

>>44249 >>44415 >Feeding sections of it into an ai to generate the scaffold or common algos. This is inferior in every way to using an API/library with static linking from trusted sources and anyone who disagrees is a retard

>>44689 > All the ai data centers will prob be retrofitted for sentiment analysis and surveillance when they finally admit no one wants it and they're hitting diminishing returns. Yeah, but they're soykaf for surveillance too.

For the grunt work of maintaining a company's legacy, poorly written code, there's really no better tool than LLMs. If you need to migrate 100 jobs from cron to autosys, you'll be happy to have access to an LLM. Many jobs don't require innovative, greenfield development. >>44689 The internet, FOSS and ad networks, has made paying for software as a concept, much less relevant to normal people than in the 90s. Specialized software already exists for every industry that needs it, and outside of tech companies, there's no desire to replace what already works.

>>44695 I'll double down on this. If some guy in the 90s who's an average coder, not an MIT wizard, magically had access to an LLM, he'd become a billionaire .

>>44695 >there's really no better tool because you read it in g or in some ai marketing brochure? have you ever actually done something like this? because I have and the last thing I would ever do would be to add a llm in the middle. it is only "grunt work" and toiling if your only problem with it is that it uses old-tool instead of new-tool. if you mechanically translate it from old-tool to new-tool, you still have a pile of garbage, but now it is using the new thing, which is probably even more inscrutable than the old thing. the point of updating such codebases to modern tooling is that it gives you an opportunity to refactor and optimize years of rushed changes and technical debt, ideally you're cutting the amount of code by >50% if you give an llm a pile of garbage (that works) written for adabas, and tell it to change it to use postgres, now you just have a pile of garbage written for postgres, but this one might or might not work, and you're going to have to inspect even more code, even more carefully, to find out >>44689 the worst part is that for some reason it is executives that are pushing it, not programmers. this soykaf doesn't happen in any other field, no healthcare executive is telling doctors which scalpels they should use or how they should do surgery. imo the problem is that they never really figured out how to measure the productivity of programmers

>>44698 I agree with your broader point, but insurance companies are able to have their people with no medical education contradict doctors when they claim procedures or other treatments are medically necessary, despite them having a vastly inferior understanding of the field compared to actual medical professionals. Should it happen, no, but there is that one even more egregious example of it happening than when it happens to idiots telling programmers how to program. That said, I think programming would look radically different as a field if programmers were free to implement ideas in whatever way they believed was best, and I agree the idea we should be using new shiny toy instead of older toy for now other reason than new shiny toy is new and shiny is pure retardation.

>>44698 >have you ever actually done something like this? Yes, and I have no regrets about it. I saved myself from hours of mind numbing bs. >the point of updating such codebases to modern tooling is that it gives you an opportunity to refactor and optimize years of rushed changes and technical debt No, that's not the point. "Run this bash script at 6:30AM every Friday" is not an invitation for your intellectual creativity to improve anything. Changing the underlying bash script is NOT part of the ask. The point is, everything else uses the new tool, so this should too. The new tool can control jobs on other machines, the old tool was entirely local. That's it. >now you just have a pile of garbage written for postgres, but this one might or might not work, and you're going to have to inspect even more code, even more carefully, to find out What are unit tests? That's how you verify if something is broken, not "inspecting" mountains of legacy code and config written by people who are long retired. If there's a problem, which should be detected by tests, then you debug it. You're a programmer, not a historian. Your example is going from Adabas, a non-relational DB, to Postgres. That's a much more interesting, error prone task, than migrating from Sybase to Postgres, relational to relational. In that case, you're just changing data types because they use different semantics for long int. Perfect job for an LLM.

>>44700 >it's good for rewriting ancient code >ok, that's false >but it is really good for this contrived and unclear hypothetical situation that I made up! chatbot tier thinking >What are unit tests? That's how you verify if something is broken wrong. before I thought you had never done this job, but now I'm thinking you have never even used tests a test failing tells you that something doesn't work, but a test passing doesn't imply that it works, matter of fact, in most cases it just means that your tests are bad and didn't catch the bugs and you start from the assumption that the ancient codebase from the 70s has tests or that it is testable at all, which is laughable. writing a test suit requires you to carefully study the code until you understand the intent behind it, and you need code designed to be tested. you write the tests for the refactored version because that's the only one that supports tests in the first place

>>44701 >contrived and unclear hypothetical situation I've done both of these tasks at my job at a bank. Nobody bitched about my results with either, which is the ultimate unit test(though I did the Sybase thing with hand-written regex, before the proliferation of LLMs). But, no, I'm making things to argue with you, some dick head on a basket weaving forum. >the assumption that the ancient codebase from the 70s The ancient codebase is from like 2005, and has been touched by countless hands over the years. Maybe that's not ancient to you, but I was born in this millennium.

>>44698 >this soykaf doesn't happen in any other field, no healthcare executive is telling doctors which scalpels they should use Pretty sure the scalpels are highly regulated and procured en mass by logistics professionals on behalf of the hospital. Nurses, paramedics, etc. have very little agency in how they perform their jobs. In fact insurance companies tell them exactly what surgery they are allowed to perform in practice.

>>44699 to add on this: if you have an MD, being an insurance claims adjuster pays like over double what all except the topmost surgeons make. This is a sign of how bad things really are...

>>44240 I'm in east europe and even as a student 15 years ago I could surely have afforded a 20$ subscription for AI. You don't need to get claude, even though it still is the best for coding. You can get one of the cheaper chinese clones. Though if you ABSOLUTELY wanted to do "agentic" then yeah I guess it's too expensive to screw around in. I only have one agentic workflow and it can cost 10-20$ to execute once (on Sol). So I'm pretty careful with using that. Though ofc, I don't pay for it, my employer does. I probably wouldn't use it if I was a student, I'd just do the same task non-agentically by breaking the task manually (or with AI design help) and then execute each step by itself. To make sure there's no waste of tokens. Aside from that, can't help but smirk at the generalizations in this thread. I'm all for pen and paper learning in school and some early years. But once you are working for real, using slow tools is just a willful handicap. Also anything past claude 4.6 is decent. It solves entire features and all you have to do is double check the code and the tests. Finally, in europe the job marked hasn't changed that much. We don't get that many migrants, especially in the poorer half :). So it's still relatively easy to get hired and it still pays considerably higher than most other jobs of similar difficulty. With a bit of luck AI stuff will plateau due to cost and we get a few more decades of easy life.

>>44786 > once you are working for real, using slow tools is just a willful handicap It makes sense if you are paid per feature. Doesn't really make sense if you have fixed salary per month. If you are paid per hour, then you can at least be flexible and say that you can do it in less hours but with higher rates because you will use agentic AI that is hella expensive.

>>44786 Didn't mean agentic use specifically. In fact I find that too wasteful for what it offers. But for normal use, it is always good. Why not do the same work faster and then just relax? I'd gladly drop my salary by 100$ per month so I can work 2-3 hours less. And unfortunately there is more. Big companies have forced churn. Some lie about it. They say "we never fire anyone" but then go "btw we're 'raising the bar' for normal performance and if you fall behind you will be an underperformer". Which is a lot worse imo. Even if you're regularly one of the good ones, it still stresses you and it's not like there's much energy left in you for that...

Since April, I'm on the project where whole webapp is vibecoded. At the beginning, I've tried to resist this, but soon I was discovered as not vibe coder, and I was forced to use it. After several months, I have following thoughts about it. At the beginning, speed gain is massive, I was able to do things, that would take me days, in mere hours, especially when I know jack soykaf about tech stack used in project. Later, when project grew, even slightest changes requires LLM munching it for at least 15 minutes, when human could do it in 2 minutes. But no human would do it, because noone understands codebase. If there is any bug or fuarrrkup, nobody knows what to do, because everything was vibecoded. After several months, I've learnt absolutely nothing about used technology and other stuff, I was just a meat proxy.

>>44808 I have an idea, but it involves something Soviet and space-telated