Interestingly I’ve learned more about languages and systems and tools I use in the last few years working with agentic coding than I did in 35 years of artisanal programming. I am still vastly superior at making decisions about systems and techniques and approaches than the agentic tools, but they are like a really really well read intern who knows a great deal of detail about errata but have very little experience. They enthusiastically make mistakes but take feedback - at least up front - even if they often forget because they don’t totally understand and haven’t internalized it.
The claim you should know everything about everything you work on is an intensely naive one. If you’ve worked on a team of more than one there’s a lot of stuff you don’t totally grok. If you work in an old code base there’s almost every bit of it that’s unfamiliar. If you work in a massive monorepo built over decades, you’re lucky if you even understand the parts everyone considers you an expert in it.
I often get the impression folks making these claims are either very junior themselves or work basically alone or on some project for 20 years. No one who works in a team or larger org can claim they know everything in their code base. No one doing agentic programming can either. But I can at least ask the agent a question and it will be able to answer it. And after reading other people’s code for most of my adult life, I absolutely can read the LLMs. The fact a machine wrote crappy code vs a human bothers me not in the least, and at least the machine will take my feedback and act on it.
This post does not make the claim that "you should know everything about everything you work on" - its making the claim that writing code and being able to read code effectively are intrinsically linked.
Agreed. I don't know anything about turning sand into transistors or assembly but do well. So I don't know my full stack either.
What is important is not being afraid to learn the rest of your system and keeping an index.
Most importantly it's about being able to spin up on anything quickly. That's how you have wide reach. Digging in when you have to, gliding high when you have to. Appropriate level for the problem at hand.
When I was in college eons ago they taught CS folks all of engineering. "When do I need to know chem-e or analog control systems?" We asked. "You won't. You just need to be able to spin up on it enough to code it and then forget it. We're providing you a strong base."
I have also seen the learning acceleration, there's a significantly increased set of techniques and technologies I have learned how to apply.
From a person perspective though, I'm apprehensive about the effect AI will have on the human "very well read intern." People who know a lot very deeply about specific areas are fascinating to talk to, but now almost everyone is able to at least emulate deep knowledge about an area through the use of AI. The productivity is there, but the human connection is missing.
"Hey! Just popping in to say that agentic coding is actually pretty great and is making me better in all the ways; but also want to say at same time that it's actually not all that different from anything else, so we can chalk up any critique to it to individual naivety and bias."
Using AI to go faster is optimizing the wrong thing. At every place I've worked, the "code writing" part takes the least amount of time, compared to all the other things you need to do in order to implement a feature. Let's examine a feature that takes a day to code:
First, you've got to plan everything, using whatever Agile or Waterfall planning ritual your company uses, that can take days or even weeks. Then you need to write a design doc with your proposed design, and get that reviewed by your peers/teammates. Again, another week for any substantial feature. If there are multiple teams involved, you need to get buy-in and design agreement among those multiple teams, let's add another week. At some places, you need approval to commence work, which can take multiple days, depending on the approver's schedule and availability.
Then, you take a day and write the code and make sure it passes tests.
Then, it's code review time, and this can involve a lot of back and forth with your team, resulting in multiple iterations and additional code reviews. Another "days or weeks" stretch. At bigger companies, you're going to need to pass all sorts of reviews from other departments, like legal, privacy, performance, accessibility... even if done in parallel, let's add a conservative 2 weeks. Finally, you push to staging, and need to get some soak time internally among dogfooders, so you have some confidence that it's working. +1 week. Then you're ready to push from staging to prod, but since you work at a serious company, nothing goes to 100% prod right away--you need to slowly ramp up and check feedback/metrics in case you need to roll back. The ramp to fully launched could take another two weeks.
So here's a feature that took, what, maybe two months from design to release, and we're falling all over ourselves to optimize the part that took a day so that it takes 5 minutes instead...
> only a skilled developer who's thinking critically, and comfortable operating at the architectural level, can spot issues in the thousands of lines of generated code, before they become a problem.
An additional factor: to find issues in generated code, the developer has to care. Many developers (especially at big firms) are already profoundly checked out from their work and are just looking for a way to close their tickets and pass the buck with the minimum possible effort. Those developers - even the capable ones - aren't going to put in the effort to understand their generated code well enough to find issues that the agents missed. Especially during the current AI-driven speed mania.
As a senior developer, 25+ years, I have been thrown recently into a meeting "hey can you join in for 5 mins". I really don't like these meetings where you're dragged in in the middle of them without any clue.
The questions came flying in fast, without any introduction, and this was about an external integration out of a dozen. They have their own lingo, different from ours, to make the situation worse.
I had a _very hard time_ making sense of the questions, as I indeed relied heavily on a model to produce these integrations (extremely boring job + external thick specs provided).
I'm still positive these would have simply not happened in a 10x the time if I did not use models, however, I'm now carefuly considering re-documenting the "ohhs" and "aahs" of these so that these kind of uncomfortable moments never happen again.
I haven't felt so clueless and embarassed in a meeting, ever. All I could say was "I'll get back to you on that one, and that one, and this one".
Cognitive debt is very real, and it hurts worse than technical debt on a personal level! Tech debt is shared across the team, cognitive debt is personal, and when you're the guy that built the thing, you should know better!
To be continued... But from now on, the work isn't done if I don't get a little 5 mins flash-card type markdown list of "what is this" and "what is that", type glossary.
What kind of place do you work where you get dragged into a meeting halfway through and then are peppered with technical questions without context, that you're expected to answer on the spot? Please let us know because I'm sure a lot of us want to avoid such a place.
"I'll need to study the docs and code to answer these questions properly" is a perfectly fine (and very diplomatic) response to treatment like that.
Not OP, but similar context (~20yr exp.). You absolutely can get away with "I'll need to dig more into this to give you a good answer" but you are _for sure_ expected to have at least some answer ready-to-go. Especially if it's under your purview.
I don’t think it was made clear: the questions were about the code op “wrote” but they used a llm so couldn’t remember any of it. Probably got there from a git blame. This happens.
This can happen more or less anywhere if you spend enough time at a company. At some point you'll get pulled into a meeting like this because others think you're in expert in a codebase/area you're not.
I kind of think this article misses the mark a little.
There is skill loss from heavy AI use.
But I want to acknowledge the awkward elephant in the room. AI Is making people too fast. I don't mean that a faster output is bad. It's a faster output and code rather than a full understanding and experience in producing the code. It's rewarding people who try to talk about business value rather than the people that are building and making safe decisions with deep knowledge.
AI: Yes, its good and it can produce some good solutions, however it ultimately doesn't know what it's doing and at the best of cases needs strong orchestrators.
We're in a cesspit of business driven development and they're not getting the right harsh and repulational punishments for bad decisions.
I've been having the same feeling too. At work, I try to do a hybrid prompt where I fill in things at method level with some placeholder pseudocode and let the prompt fill in the blanks. This helps with remembering and keeping a memory map. But its a lost cause keeping up with other's PRs that are often very verbose and high volume. For a lot of backend programming without a very complex domain, I think this works fine.
But I still want to be in touch with coding by hand and have ventured into systems programming, outside of work, which I feel AI is less useful for currently.
If you are coding by hand like the old days you are probably not literally writing everything from scratch anyway, you are copy pasting a bunch of shit off google and stackoverflow or installing open source libraries.
I also reuse a lot of my own code. Either from libraries I built or just directly copy pasting (like boilerplate code for setting up the basics of something in my style).
Actually ignore my comment I misunderstood the premise. I meant not vibe coding is the way to save time with production issues. Not the other way around!
AI doesn't automatically make us better human beings, but they only expose our worst parts. Most people are not born great leaders and managers (need rigorous training and experience), and empowering them with AI kind of pushes them into a spot where they suddenly need to "lead". To fight brainrot from AI overuse, we must try harder to maintain that developer's priority list.
it's a fairly new way of doing things. I predict, in the future it will be more formalized and standardized like AGILE and SCRUM and all that boring stuff.
The result of that though would be establishment of development patterns that are good practices.
The rule of thumb is: An agent can write it, but a human has to understand it before it gets pushed to prod.
I'm still not convinced about the doom and gloom over developers being replaced. I'm not a dev as part of my main job function, but where I do use LLMs, it has been to do things I couldn't have done before because I just didn't have time, and had to de-prioritize. You can ship more and better features. I think LLMs being tools and all, there is too much focus on how the tool should be used without considering desired and actualized results.
If you just want an app shipped with little hassle and that's it, just let Claude do most of the work and get it over with. If you have other requirements, well that's where the best practices and standards would come in the future (I hope), but for now we're all just reading random blog posts and see how others are faring and experimenting.
> The rule of thumb is: An agent can write it, but a human has to understand it before it gets pushed to prod.
The article essentially claims that no, that line of thinking is false. If the agent writes all of it (or too much of it, where "too much" is still not well defined), then your ability to understand it will atrophy with time, and you will either a) never push to prod, because you can't understand it well enough, or b) push to prod anyway, and cause bugs and outages.
I think the article is correct.
> I'm still not convinced about the doom and gloom over developers being replaced.
Agreed. The agents are just not good enough to write code unsupervised, or supervised by people without senior-level skills. And frankly it's hard to imagine them getting there. Each new release of the coding tools/models is a mixed bag. Some things are better, some things are worse, and the gains are diminishing with each iteration. I am afraid that we're going to hit a ceiling at some point, at least with the transformer architecture.
> but for now we're all just reading random blog posts and see how others are faring and experimenting.
Yes, exactly, and many people are not faring well. The article cites several examples of people feeling less capable after using LLMs to write code for a while.
I've been using AI tools to brainstorm approaches and sometimes generate code, but actually doing the typing myself. That way I'm less likely to forget the mechanics and programming language over time.
Same. Most of what I do is ask for an implementation plan, with minimal code, or no code, or pseudocode, and then write the actual code myself. This is for open source work, where the entire point of my enjoyment is that I write the code myself. I honestly wouldn't bother being an open source maintainer if the entire thing was just prompting an LLM to write code, and then reviewing it. That doesn't sound fulfilling at all.
If this was an actual paid job, I do wonder how that would change my LLM use. The reason I'm a software developer at all is because I love the craft. The act of building, of using my brain to transform ideas into code... that's what I enjoy. If it was just prompting an LLM, would I still do that job? I don't know. I'd probably start looking into the idea of switching careers, at least.
One approach you can use is to ask it to never write the code for you, which forces it to explain and then once you try the idea by coding yourself you get a better understanding of it. I use this approach with code I am required to maintain. It still bites me sometimes because the models still mixes a lot of incorrect information (usually just stuff that was correct in the past but is incorrect now). For throwaway and easy to verify scripts I ask it to generate, but I do ask to avoid over engineering and trying to catch all corner cases cause in scripts I prefer just letting things error as they are better understood as a step that failed. I also avoid languages I find hard to read (like powershell) and prefer to generate things that are short to fit in the monitor so I can read everything and understand (python, bash, batch are my goto scripting languages).
Same. I've also configured the system prompt to never give me a full solution or write a code for me. So whenever I ask it a question it produces a short 10 line example or even a pseudocode. This is far easier for me to reason about.
I still reject > 50% of AI suggestions, because they're too mediocre, like moving code for no reason or sometimes it is just plain wrong.
Remembering and understanding aren't the same thing. Merely reciting facts doesn't automatically give you the ability to apply those facts to solve problems.
Me too, ... more or less. I'm mostly still typing, sometimes copy-and-pasting with typed changes, and rarely copy-and-pasting verbatim. With the caveat that in some cases, like prototypes, proofs-of-concept, and porting code between languages; then maybe many lines are copy-and-pasted verbatim.
I mean that’s been my line every time someone makes impressed noises when I say I’m a programmer - it’s really not that hard, it’s really just a question of whether you like it enough to put the work in, like anything else. “Don’t you have to be a math wiz?” No dude 95% of the time whatever you’re trying to do already has a very well researched approach, a lot of times you’re just picking which pre-vetted solution to adapt to your needs.
Right. Like anywhere the conceptual problems haven't been all figured out yet, or where higher order effects happen with scale or particular shapes of data/substrate and you don't know them in advance.
Sometimes hard like interesting and you get to do really novel thinking. A load of p2p/decentralised things are hard like this.
Also sometimes hard like you get to a particular challenge and it turns out to be a notoriously unsolved mathematical thing, or you push against subtle boundaries of core libraries, runtimes, systems etc. Working with metagenome assemblies is this kind of hard.
Honestly the hard code I've done made such a difference to my brain. There's plenty of trivial stuff I'm happy to have automated, but of I can't work on the hard problems I may as well not be involved at all.
Yeah this is the same conclusion I have. I primarily use AI for UI code, and guess what, it's all basically mechanical drudgery anyways. Put a div here, or put a Box here, apply some style rules, etc. This shit should have been automated decades ago yet for some reason we're still writing the same stuff with a different "twist" today.
Now if your career is built on writing out the same boilerplate code in its infinite slight variations every day, congrats, you've been automated. Thank god we can free up our intellects to focus on the actual hard problems, the ones that are somewhat cutting edge, the ones that actually push our field and humanity forward.
Literally every example of AI generated code (without significant human input) is just basic stuff that is wholly unimpressive. Oh wow, you had an AI generate a Next.js app? It's writing HTML for you? It made a generic SAAS? Guess I'll become a farmer now.
Or, wait, I'll continue to write my multithreaded real-time multiplayer network for a MMO, since the AI currently generates something that would get me fired 10 seconds ago if I tried to push it to production.
It's amazing how you introduce just the slightest difficulty or novelty to an AI and it just craps the bed. And then you go online and apparently we're gonna be replaced -6 months ago or something.
I think of it as driver's seat vs back seat vs passenger seat. You always take the back seat and eventually you will forget how to drive. You insist on always being in the front seat and you will miss out on the occasions where the LLM happens to know the area very well, like working with an unfamiliar library or problem domain. If it is a place that you are just passing through, it's a great to let it take the wheel and see where it will takes you. If it is a place that you need to become familiar with, it's great to have a dependable navigator beside you.
My sense is that a decade from now, the people who generally see their place as the driver seat but recognize when its not are going to be writing the code that matters.
I tend to think of it as like two pilots as on commercial airliners: you always have one pilot flying and one pilot monitoring.
You can debate with agentic coding who is monitoring and who is flying but, if we assume the user is monitoring what that means, in practice, for me is that I'm reading and making sure I understand all the changes the agent is proposing to make, as well as providing instruction, guidance, correction, etc. That includes reading and understanding all the code changes.
the funny thing is once the llms got mostly good enough in november 2025 for me, it was mind boggling how much it helped me get stuff out of my head with ease.
its easier for me to code now, because its like i have a 24/7 insane intern that needs to be supervised via pair programming but also understands most topics enough to be useful/ dangerous.
ironically ive been spending much of my time iterating on ways to improve model reasoning and reliability and aside from the challenge of benchmark design, ive had some pretty good success!!
my fork of omp: https://github.com/cartazio/oh-punkin-pi has a bunch of my ideas layered on top. ultimately its just a bridge till i’ve finished the build of the proper 2nd gen harness with some other really cool stuff folded in. not sure if theres a bizop in a hosted version of what ive got planned, but the changes ive done in my forks have made enough difference that i can see the different in per model reasoning
>and then pulls the slot machine lever over and over
Does anyone really do this? You want verification and self-correction in a loop, not rerolling and cherrypicking. The non-determinism point is really tiresome to hear over and over.
There’s too much in this article to comment on it all, but if we zoom into the first claim:
> An increase in the complexity of the surrounding systems to mitigate the increased ambiguity of AI's non-determinism.
My question is why isn’t there an effort from the author to mitigate the insane things that LLMs do? For example, I set up a hexagonal design pattern for our backend. Claude Code printed out directionally ok but actually nonsensical code when I asked it to riff off the canonical example.
Then, I built linters specific to the conventions I want. For example, all hexagonal features share the same directory structure, and the port.py file has a Protocol class suffixed with “Port”.
That was better but there was a bunch of wheel spinning so then I built a scaffolder as part of the linter to print out templated code depending on what I want to do.
Then I was worried it was hallucinating the data, so I wrote a fixture generator that reads from our db and creates accurate fixtures for our adapters.
Since good code has never been “explained for itself 100%, without comments”, I employ BDD so the LLM can print out in a human readable way what the expected logical flow is. And for example, any disable of a custom rule I wrote requires and explanation of why as a comment.
Meanwhile, I’m collecting feedback from the agents along the way where they get tripped up, and what can improve in the architecture so we can promote more trust to the output. Like, I only have a fixture printer because it called out that real data (redacted yes) would be a better truth than any mocks I made.
Finally, code review is now less focused on the boilerplate and much more control flow in the use_case.
The stakes to have shitty code in these in-house tools is almost zero since new rules and rule version bumps are enforced w a ratchet pattern. Let the world fail on first pass.
Anyway, it seems to me like with investment you can slap rails on your code and stay sharp along the way. I have a strong vision for what works, am able to prove it deterministically with my homespun linters, and am being challenged by the LLMs daily with new ideas to bolt on.
So I don’t know, seems like the issue comes down to choosing to mistrust instead of slap on rails.
Edit: I wanted to ask if anyone is taking this approach or something similar, or have thought about things like writing linters for popular packages that would encourage a canonical implementation (I have seen some crazy crazy modeling with ORMs just from folks not reading the docs). HMU would love to chat youngii.jc@gmail
Lars we are on the same page. I use LLMs to help me scope and get a second set of eyes on the high levels of a task. Then I write the code. Often I automate boilerplate or boring objects but sometimes it's faster/better for me to just write them. Then I will ask an LLM to say write some tests. Then I will focus on the cases they missed and write those myself.
I have been described as a decel and a Luddite though so be weary of my opinions.
Re vendor lock in point: this is a harness issue really. Sure, CC is restricted to Anthropic models, but it's not the only harness out there. So if one vendor has an outage or botches the quality of their models due to compute shortage, you can switch to another vendor. LLMs are the easiest to switch. Of course, if hardware costs go up, so will all AI vendors. The only way out for the employer would be to directly buy the hardware (or do a fixed price deal with a cloud provider).
Re the understanding code point: you can still use LLMs to understand code. If you write the spec without knowing anything about the code, of course the architecture might suck. Maybe there is already a subsystem that you can modify and extend instead of adding a completely new one for the new feature you are adding, etc.
I use LLMs for my daily workflows and they do understand code perfectly and much more quickly than if I read it.
CC isn’t even limited to Anthropic models, there’s a post on the front page right now to use it with Deepseek V4 since Deepseek provides an Anthropic compatible API and CC reads API URLs from env variables so you can override them.
I’ve build a configuration transpiler to Claude code and codex and found I can switch pretty quickly between both and run both at once. At the moment codex performs better. Prior CC did. There is no vendor lockin and this is an old canard in technology that LLMs in fact themselves make irrelevant. Once you’ve got an implementation that uses X converting it to Y is almost trivial with an LLM because the spec is canonical in the reference.
Would like to see a study of brain scans during flow, manual programming, compared to code review. If the conclusion is different parts of the brain are activated, then orchestration is a separate activity entirely. Reading code is not the same as writing code.
However, the code review study needs to compare between surface scanning and reviewing long enough to get over a theoretical slough of perspective: when you assume the coding chair and are in their frame, whether the brain shifts into a different cognitive mode.
Otherwise, just stamping "Looks good to me" is likely to lead to the same atrophy. There's no critical thought, even a self-summary of the change or active questioning.
Thoughtful, deliberate code review just plain takes longer. AI can help here a lot, although it still takes over the "get into review mode" process.
I absolutely feel like a "different" part of my mind is loaded when seriously engineering something myself vs vibecoding+reviewing. Even the reviewing is more annoying in the latter mental context.
Code review alone is kind of like being able to understand a foreign language enough to read it, but not really understand it in flowing conversation or being able to speak it, much less construct a complex piece of literature.
Retention also suffers, as you will quickly forget what you just reviewed. What is the last PR you remember?
This is how I feel about things. Its like someone is demanding that I become a manager, when I was perfectly happy being a IC. And now I have to figure out how to be a manager of AI agents while at the same time not lose my ability to judge their work, or plan effectively, even though I'm not supposed to be doing things "by hand" anymore. But doing things "by hand" is how I reasoned through problems and figured out the plan to begin with.
I can’t say that I’ve felt my skills atrophy, but I’ve also never found backend web development to be that difficult. 90% of my job for my entire career could be described as digital plumber.
Software engineers and their leadership have been pushing back on terrible product managers for decades. AI isn't the reason to stop our time honored tradition. If anything we can write the emails faster now
Only way to cope with this I found is to grind leetcode or advent of code. It's kinda funny how fast this all changed. Less funny part is the fact that I'm now kinda feared for my job in some time.
This is exactly the same problem of "mechanical engineers' job is to design parts, not machine them, so we'll take training on machines out of the mech eng curriculum." Result: fresh mech eng grads do not know how to properly design parts because they have no idea how they are machined.
Doesn’t matter, if you ask for physically impossible features to cut this is something you could technically do. Or you ask for a feature that adds multiple setups to an otherwise simple part and makes it wildly expensive
The CNC doesn't know either. What usually happens is that an engineer at the CNC shop figures it out for you.
Knowing some machining still lets you design parts and assemblies that are some combination of cheaper, better, etc. This is noticeable with precision or high performance assemblies. And how many revisions are needed.
How can we solve this at a more fundamental level?
I think many people already recognize the problem:
-“Our ability to write code is being damaged.”
-“If our ability to write code declines, our ability to recognize good code also declines.”
But the problem is that the market no longer works without LLMs.
Freelance rates and deadlines are now calibrated around LLM-assisted output. Even clients who write “do not vibe code” often set deadlines that are impossible to meet unless you use something like vibe coding. The client’s expectations themselves are becoming abnormal.
That is the irony of the market.
I honestly do not know what to do.
Recent Hacker News discussions are mostly a negative echo chamber about AI use. In other places, it is often the opposite: only positive echo. But almost nobody discusses the actual solution.
The main topics I keep seeing are roughly these:
1. Is the large repository PR system failing a fundamental stress test? Or should AI-generated(GEN AI) code simply not be merged? If PR review is moving from handmade production to mass production, how should the PR system change? Or should it remain the same?
2. As vendor lock-in continues, can we move toward local LLMs to escape it? Are cost and harness design manageable? What level of local model is required to reach a similar coding speed?
3. If we are forced to use agentic coding, how do we avoid damaging our own ability to code?
There is a passage from Christopher Alexander that I keep thinking about:
“A whole academic field has grown up around the idea of ‘design methods’—and I have been hailed as one of the leading exponents of these so-called design methods. I am very sorry that this has happened, and want to state, publicly, that I reject the whole idea of design methods as a subject of study, since I think it is absurd to separate the study of designing from the practice of design. In fact, people who study design methods without also practicing design are almost always frustrated designers who have no sap in them, who have lost, or never had, the urge to shape things.”
— Christopher Alexander, 1971
This quote feels relevant to programming now.
If we separate the study and supervision fo programming from the actual practice of making, something important may be lost.
In architecture, there is this idea that without practice, the architect loses meaning. But now the market is forcing the separation.
People with enough symbolic capital and high status have the freedom not to use AI. But people lower in the market are under pressure to use it.
So I think the discussion now needs to move beyond whether AI coding is good or bad.
The real question is How do we keep using AI because the market demands it, while still preserving the human practice that makes programming meaningful and keeps our judgment alive?
I think these are the important question.
How do you maintain market value without using AI?
Or, if you do use AI, how do you avoid being treated as low-quality?
If you do not use AI, how can you remain more competitive than people who do use it?
If you do use AI, what advanatge do you have over people who do not use it, and how should you position yourself?
I know that agentic coding can cause skill degradation. I can feel it happening to me already. But for someone like me, who does not have strong status, credentials, or symbolic capital, social and market pressure makes AI almost unavoidable.
What frustrates me is that I do not see practical answers anywhere.
I use AI mainly to improve the parts which I know I am pretty bad and need help and restrict it as I gain competency. Which means choosing to say no to copy pasting most things I wish to exercise my brain for and reading, understanding, processing and writing it more diligently. Treating brain like any other muscle that requires a gym, we do have cars and segways that can help us travel long distances quickly but we still require as a specie to walk and run. People with disability and underprivileged backgrounds also need these tools so it is a good option to have generally. Secondly we are human, we have made very good tools, we need to stop obsessing with being more productive in output and learn to be better, more considerate and respectful to each other and other species to make this planet generally for better world. It is the only one we got. At the same time, be more mindful about where and how we utilise our time and attention, they are very limited and the most valuable resources we can possibly possess.
"How can we solve this at a more fundamental level?"
Stop using AI for coding. Period...there is no other solution. You can't make it work, nobody else can either. Without determinism, the entire process is useless. We need to stop trying to act like we all know that this isn't true. We have given it a chance, it failed, time to move on to something else no matter how much the VCs and execs don't want to. Those that do move on have a chance, the others have no future in software.
The issue is that you will end up without a job if the trend continues. It's similar to many cases of technical innovation - you can still have a few workers who do handcrafted works, but most of them have to use the machines, that may produce work of inferior quality but at much higher speed.
The market realigns, and unless you handwrite the highest possible quality at a quick pace, you won't be competitive with the vibe-coders who can fix a hundred issues a month.
It was the same with gps-assisted driving, now most people can't orient themselves autonomously. Worse, there are no roadsigns with directions installed, meaning that you are stuck with using the GPS.
"unless you handwrite the highest possible quality at a quick pace"
That's exactly what I do. I know I am lucky to be gifted in this skillset. But that's not a good reason to excuse people destroying the market for everyone.
I agree with what you are saying, but if I cannot get work, I may literally have nothing to eat tomorrow.
So while I agree with your point, it does not feel like a practical answer for my situation. For someone who is already well known and has enough reputation, refusing to use AI may be a matter of principle. But I am dealing with survival.
I do not think your answer is bad. But because this is a survival problem, it is difficult for me to risk everything on principle.
In other words, I know that your answer may be the morally correct one. If everyone boycotted this, perhaps it would not be adopted so aggressively.
But I cannot do that.
What I need is a way to use AI while degrading my own ability as little as possible, and while still preserving my skills.
I am not saying you are wrong. I am saying that your answer is too idealistic for someone in my position.
I'm not being idealistic. I'm being very practical. You have the survival problem exactly backwards. Continuing to use it, that's the real danger in a practical sense. That only leads one direction and that direction isn't in your best interest.
Nope. 1) Skills don't go away, you just get better at the things you do regularly, but your still have your old skills, 2) You only have vendor lock-in if you use lock-in devices (stop using Claude Code), 3) It's not an increase in complexity, it's a replacement, in order to gain efficiency (see: the cotton gin), 4) The increased cost is negligible considering average salary and resulting productivity
Im seeing the word "agentic" a lot here. Is there a difference between "Agentic Coding" and "I put prompt into gpt or claude and pasted code into my file" ?
I think ignoring all else, generating code is not a new layer of abstraction. It's the same abstraction, we just have codegen machines now. The same skills are important regardless if the person is typing in the code or if a machine is producing it.
The claim you should know everything about everything you work on is an intensely naive one. If you’ve worked on a team of more than one there’s a lot of stuff you don’t totally grok. If you work in an old code base there’s almost every bit of it that’s unfamiliar. If you work in a massive monorepo built over decades, you’re lucky if you even understand the parts everyone considers you an expert in it.
I often get the impression folks making these claims are either very junior themselves or work basically alone or on some project for 20 years. No one who works in a team or larger org can claim they know everything in their code base. No one doing agentic programming can either. But I can at least ask the agent a question and it will be able to answer it. And after reading other people’s code for most of my adult life, I absolutely can read the LLMs. The fact a machine wrote crappy code vs a human bothers me not in the least, and at least the machine will take my feedback and act on it.
What is important is not being afraid to learn the rest of your system and keeping an index.
Most importantly it's about being able to spin up on anything quickly. That's how you have wide reach. Digging in when you have to, gliding high when you have to. Appropriate level for the problem at hand.
When I was in college eons ago they taught CS folks all of engineering. "When do I need to know chem-e or analog control systems?" We asked. "You won't. You just need to be able to spin up on it enough to code it and then forget it. We're providing you a strong base."
That holds even within just large code bases.
From a person perspective though, I'm apprehensive about the effect AI will have on the human "very well read intern." People who know a lot very deeply about specific areas are fascinating to talk to, but now almost everyone is able to at least emulate deep knowledge about an area through the use of AI. The productivity is there, but the human connection is missing.
Nothing in the article made that claim.
First, you've got to plan everything, using whatever Agile or Waterfall planning ritual your company uses, that can take days or even weeks. Then you need to write a design doc with your proposed design, and get that reviewed by your peers/teammates. Again, another week for any substantial feature. If there are multiple teams involved, you need to get buy-in and design agreement among those multiple teams, let's add another week. At some places, you need approval to commence work, which can take multiple days, depending on the approver's schedule and availability.
Then, you take a day and write the code and make sure it passes tests.
Then, it's code review time, and this can involve a lot of back and forth with your team, resulting in multiple iterations and additional code reviews. Another "days or weeks" stretch. At bigger companies, you're going to need to pass all sorts of reviews from other departments, like legal, privacy, performance, accessibility... even if done in parallel, let's add a conservative 2 weeks. Finally, you push to staging, and need to get some soak time internally among dogfooders, so you have some confidence that it's working. +1 week. Then you're ready to push from staging to prod, but since you work at a serious company, nothing goes to 100% prod right away--you need to slowly ramp up and check feedback/metrics in case you need to roll back. The ramp to fully launched could take another two weeks.
So here's a feature that took, what, maybe two months from design to release, and we're falling all over ourselves to optimize the part that took a day so that it takes 5 minutes instead...
An additional factor: to find issues in generated code, the developer has to care. Many developers (especially at big firms) are already profoundly checked out from their work and are just looking for a way to close their tickets and pass the buck with the minimum possible effort. Those developers - even the capable ones - aren't going to put in the effort to understand their generated code well enough to find issues that the agents missed. Especially during the current AI-driven speed mania.
The questions came flying in fast, without any introduction, and this was about an external integration out of a dozen. They have their own lingo, different from ours, to make the situation worse.
I had a _very hard time_ making sense of the questions, as I indeed relied heavily on a model to produce these integrations (extremely boring job + external thick specs provided).
I'm still positive these would have simply not happened in a 10x the time if I did not use models, however, I'm now carefuly considering re-documenting the "ohhs" and "aahs" of these so that these kind of uncomfortable moments never happen again.
I haven't felt so clueless and embarassed in a meeting, ever. All I could say was "I'll get back to you on that one, and that one, and this one".
Cognitive debt is very real, and it hurts worse than technical debt on a personal level! Tech debt is shared across the team, cognitive debt is personal, and when you're the guy that built the thing, you should know better!
To be continued... But from now on, the work isn't done if I don't get a little 5 mins flash-card type markdown list of "what is this" and "what is that", type glossary.
"I'll need to study the docs and code to answer these questions properly" is a perfectly fine (and very diplomatic) response to treatment like that.
It's quite common to search for the author of a piece of code to ask questions about that code.
There is skill loss from heavy AI use.
But I want to acknowledge the awkward elephant in the room. AI Is making people too fast. I don't mean that a faster output is bad. It's a faster output and code rather than a full understanding and experience in producing the code. It's rewarding people who try to talk about business value rather than the people that are building and making safe decisions with deep knowledge.
AI: Yes, its good and it can produce some good solutions, however it ultimately doesn't know what it's doing and at the best of cases needs strong orchestrators.
We're in a cesspit of business driven development and they're not getting the right harsh and repulational punishments for bad decisions.
But I still want to be in touch with coding by hand and have ventured into systems programming, outside of work, which I feel AI is less useful for currently.
Nothing stopping you from iterating with the agent till the code is the exact same quality that you yourself would write
I don't want my code quality, I want AGI code quality - that's what I was promised and jetpacks and flying cars too!
There are some times after iterating so much I’m not sure if I’ve even saved time because going from “it works” to “it’s up to quality” takes so long
And yea usually does for me
If you are coding by hand like the old days you are probably not literally writing everything from scratch anyway, you are copy pasting a bunch of shit off google and stackoverflow or installing open source libraries.
Wait, is this the same AWS I have been using?
The result of that though would be establishment of development patterns that are good practices.
The rule of thumb is: An agent can write it, but a human has to understand it before it gets pushed to prod.
I'm still not convinced about the doom and gloom over developers being replaced. I'm not a dev as part of my main job function, but where I do use LLMs, it has been to do things I couldn't have done before because I just didn't have time, and had to de-prioritize. You can ship more and better features. I think LLMs being tools and all, there is too much focus on how the tool should be used without considering desired and actualized results.
If you just want an app shipped with little hassle and that's it, just let Claude do most of the work and get it over with. If you have other requirements, well that's where the best practices and standards would come in the future (I hope), but for now we're all just reading random blog posts and see how others are faring and experimenting.
Yeah, likely
> development patterns that are good practices.
Wait, now you lost me
The article essentially claims that no, that line of thinking is false. If the agent writes all of it (or too much of it, where "too much" is still not well defined), then your ability to understand it will atrophy with time, and you will either a) never push to prod, because you can't understand it well enough, or b) push to prod anyway, and cause bugs and outages.
I think the article is correct.
> I'm still not convinced about the doom and gloom over developers being replaced.
Agreed. The agents are just not good enough to write code unsupervised, or supervised by people without senior-level skills. And frankly it's hard to imagine them getting there. Each new release of the coding tools/models is a mixed bag. Some things are better, some things are worse, and the gains are diminishing with each iteration. I am afraid that we're going to hit a ceiling at some point, at least with the transformer architecture.
> but for now we're all just reading random blog posts and see how others are faring and experimenting.
Yes, exactly, and many people are not faring well. The article cites several examples of people feeling less capable after using LLMs to write code for a while.
Also, let’s not forget. The developer is rarely the person pitching the feature, and is normally given the constraints and the PRD…
Soooo people can keep tiptapping on the keyboard, but eventually they need to open their mind to the possibility that “the old way” is actually dead.
If this was an actual paid job, I do wonder how that would change my LLM use. The reason I'm a software developer at all is because I love the craft. The act of building, of using my brain to transform ideas into code... that's what I enjoy. If it was just prompting an LLM, would I still do that job? I don't know. I'd probably start looking into the idea of switching careers, at least.
I still reject > 50% of AI suggestions, because they're too mediocre, like moving code for no reason or sometimes it is just plain wrong.
Sometimes hard like interesting and you get to do really novel thinking. A load of p2p/decentralised things are hard like this.
Also sometimes hard like you get to a particular challenge and it turns out to be a notoriously unsolved mathematical thing, or you push against subtle boundaries of core libraries, runtimes, systems etc. Working with metagenome assemblies is this kind of hard.
Honestly the hard code I've done made such a difference to my brain. There's plenty of trivial stuff I'm happy to have automated, but of I can't work on the hard problems I may as well not be involved at all.
Now if your career is built on writing out the same boilerplate code in its infinite slight variations every day, congrats, you've been automated. Thank god we can free up our intellects to focus on the actual hard problems, the ones that are somewhat cutting edge, the ones that actually push our field and humanity forward.
Literally every example of AI generated code (without significant human input) is just basic stuff that is wholly unimpressive. Oh wow, you had an AI generate a Next.js app? It's writing HTML for you? It made a generic SAAS? Guess I'll become a farmer now.
Or, wait, I'll continue to write my multithreaded real-time multiplayer network for a MMO, since the AI currently generates something that would get me fired 10 seconds ago if I tried to push it to production.
It's amazing how you introduce just the slightest difficulty or novelty to an AI and it just craps the bed. And then you go online and apparently we're gonna be replaced -6 months ago or something.
People need a reality check.
My sense is that a decade from now, the people who generally see their place as the driver seat but recognize when its not are going to be writing the code that matters.
You can debate with agentic coding who is monitoring and who is flying but, if we assume the user is monitoring what that means, in practice, for me is that I'm reading and making sure I understand all the changes the agent is proposing to make, as well as providing instruction, guidance, correction, etc. That includes reading and understanding all the code changes.
its easier for me to code now, because its like i have a 24/7 insane intern that needs to be supervised via pair programming but also understands most topics enough to be useful/ dangerous.
ironically ive been spending much of my time iterating on ways to improve model reasoning and reliability and aside from the challenge of benchmark design, ive had some pretty good success!!
my fork of omp: https://github.com/cartazio/oh-punkin-pi has a bunch of my ideas layered on top. ultimately its just a bridge till i’ve finished the build of the proper 2nd gen harness with some other really cool stuff folded in. not sure if theres a bizop in a hosted version of what ive got planned, but the changes ive done in my forks have made enough difference that i can see the different in per model reasoning
Does anyone really do this? You want verification and self-correction in a loop, not rerolling and cherrypicking. The non-determinism point is really tiresome to hear over and over.
Yes, lots of people. It’s a whole issue.
> An increase in the complexity of the surrounding systems to mitigate the increased ambiguity of AI's non-determinism.
My question is why isn’t there an effort from the author to mitigate the insane things that LLMs do? For example, I set up a hexagonal design pattern for our backend. Claude Code printed out directionally ok but actually nonsensical code when I asked it to riff off the canonical example.
Then, I built linters specific to the conventions I want. For example, all hexagonal features share the same directory structure, and the port.py file has a Protocol class suffixed with “Port”.
That was better but there was a bunch of wheel spinning so then I built a scaffolder as part of the linter to print out templated code depending on what I want to do.
Then I was worried it was hallucinating the data, so I wrote a fixture generator that reads from our db and creates accurate fixtures for our adapters.
Since good code has never been “explained for itself 100%, without comments”, I employ BDD so the LLM can print out in a human readable way what the expected logical flow is. And for example, any disable of a custom rule I wrote requires and explanation of why as a comment.
Meanwhile, I’m collecting feedback from the agents along the way where they get tripped up, and what can improve in the architecture so we can promote more trust to the output. Like, I only have a fixture printer because it called out that real data (redacted yes) would be a better truth than any mocks I made.
Finally, code review is now less focused on the boilerplate and much more control flow in the use_case.
The stakes to have shitty code in these in-house tools is almost zero since new rules and rule version bumps are enforced w a ratchet pattern. Let the world fail on first pass.
Anyway, it seems to me like with investment you can slap rails on your code and stay sharp along the way. I have a strong vision for what works, am able to prove it deterministically with my homespun linters, and am being challenged by the LLMs daily with new ideas to bolt on.
So I don’t know, seems like the issue comes down to choosing to mistrust instead of slap on rails.
Edit: I wanted to ask if anyone is taking this approach or something similar, or have thought about things like writing linters for popular packages that would encourage a canonical implementation (I have seen some crazy crazy modeling with ORMs just from folks not reading the docs). HMU would love to chat youngii.jc@gmail
I have been described as a decel and a Luddite though so be weary of my opinions.
Re the understanding code point: you can still use LLMs to understand code. If you write the spec without knowing anything about the code, of course the architecture might suck. Maybe there is already a subsystem that you can modify and extend instead of adding a completely new one for the new feature you are adding, etc.
I use LLMs for my daily workflows and they do understand code perfectly and much more quickly than if I read it.
Why make this assumption so confidently?
The arrival of the electronic computer did not turn human computers into programmers, it simply eliminated them en masse.
However, the code review study needs to compare between surface scanning and reviewing long enough to get over a theoretical slough of perspective: when you assume the coding chair and are in their frame, whether the brain shifts into a different cognitive mode.
Otherwise, just stamping "Looks good to me" is likely to lead to the same atrophy. There's no critical thought, even a self-summary of the change or active questioning.
Thoughtful, deliberate code review just plain takes longer. AI can help here a lot, although it still takes over the "get into review mode" process.
And they will deserve it.
Code review alone is kind of like being able to understand a foreign language enough to read it, but not really understand it in flowing conversation or being able to speak it, much less construct a complex piece of literature.
Retention also suffers, as you will quickly forget what you just reviewed. What is the last PR you remember?
Knowing some machining still lets you design parts and assemblies that are some combination of cheaper, better, etc. This is noticeable with precision or high performance assemblies. And how many revisions are needed.
I think many people already recognize the problem:
-“Our ability to write code is being damaged.” -“If our ability to write code declines, our ability to recognize good code also declines.”
But the problem is that the market no longer works without LLMs.
Freelance rates and deadlines are now calibrated around LLM-assisted output. Even clients who write “do not vibe code” often set deadlines that are impossible to meet unless you use something like vibe coding. The client’s expectations themselves are becoming abnormal.
That is the irony of the market.
I honestly do not know what to do.
Recent Hacker News discussions are mostly a negative echo chamber about AI use. In other places, it is often the opposite: only positive echo. But almost nobody discusses the actual solution.
The main topics I keep seeing are roughly these:
1. Is the large repository PR system failing a fundamental stress test? Or should AI-generated(GEN AI) code simply not be merged? If PR review is moving from handmade production to mass production, how should the PR system change? Or should it remain the same?
2. As vendor lock-in continues, can we move toward local LLMs to escape it? Are cost and harness design manageable? What level of local model is required to reach a similar coding speed?
3. If we are forced to use agentic coding, how do we avoid damaging our own ability to code? There is a passage from Christopher Alexander that I keep thinking about:
“A whole academic field has grown up around the idea of ‘design methods’—and I have been hailed as one of the leading exponents of these so-called design methods. I am very sorry that this has happened, and want to state, publicly, that I reject the whole idea of design methods as a subject of study, since I think it is absurd to separate the study of designing from the practice of design. In fact, people who study design methods without also practicing design are almost always frustrated designers who have no sap in them, who have lost, or never had, the urge to shape things.” — Christopher Alexander, 1971
This quote feels relevant to programming now. If we separate the study and supervision fo programming from the actual practice of making, something important may be lost.
In architecture, there is this idea that without practice, the architect loses meaning. But now the market is forcing the separation.
People with enough symbolic capital and high status have the freedom not to use AI. But people lower in the market are under pressure to use it.
So I think the discussion now needs to move beyond whether AI coding is good or bad.
The real question is How do we keep using AI because the market demands it, while still preserving the human practice that makes programming meaningful and keeps our judgment alive?
I think these are the important question. How do you maintain market value without using AI?
Or, if you do use AI, how do you avoid being treated as low-quality?
If you do not use AI, how can you remain more competitive than people who do use it?
If you do use AI, what advanatge do you have over people who do not use it, and how should you position yourself?
I know that agentic coding can cause skill degradation. I can feel it happening to me already. But for someone like me, who does not have strong status, credentials, or symbolic capital, social and market pressure makes AI almost unavoidable.
What frustrates me is that I do not see practical answers anywhere.
Stop using AI for coding. Period...there is no other solution. You can't make it work, nobody else can either. Without determinism, the entire process is useless. We need to stop trying to act like we all know that this isn't true. We have given it a chance, it failed, time to move on to something else no matter how much the VCs and execs don't want to. Those that do move on have a chance, the others have no future in software.
The market realigns, and unless you handwrite the highest possible quality at a quick pace, you won't be competitive with the vibe-coders who can fix a hundred issues a month.
It was the same with gps-assisted driving, now most people can't orient themselves autonomously. Worse, there are no roadsigns with directions installed, meaning that you are stuck with using the GPS.
That's exactly what I do. I know I am lucky to be gifted in this skillset. But that's not a good reason to excuse people destroying the market for everyone.
So while I agree with your point, it does not feel like a practical answer for my situation. For someone who is already well known and has enough reputation, refusing to use AI may be a matter of principle. But I am dealing with survival.
I do not think your answer is bad. But because this is a survival problem, it is difficult for me to risk everything on principle.
In other words, I know that your answer may be the morally correct one. If everyone boycotted this, perhaps it would not be adopted so aggressively.
But I cannot do that.
What I need is a way to use AI while degrading my own ability as little as possible, and while still preserving my skills.
I am not saying you are wrong. I am saying that your answer is too idealistic for someone in my position.