Everyone I have talked to who uses coding agents at all now uses them to write almost all their code.
The two people I know who don’t use coding agents work in government, and in a data science company working with government.
I’d say if you work at a tech company and agents aren’t writing the majority of your code, that is weird. But if you work at a traditional company that doesn’t have Claude or Codex subscriptions, there it might be pretty normal to still be writing code by hand.
Well in my personal projects I write all code by hand because people relying on agents will have forgotten how to do their fucking job in 3 years and then it will be appreciated to have people understanding what they're actually doing. That being said, in my job ai writes the code just because it's way faster.
Your "fucking job in 3 years" will be very different, and if you don't learn how to do it in time, you won't still have a job in the computer industry.
Pick an industry where you can stop learning new things once you're out of school, because the computer industry is not one of them and never has been.
My fucking job used to be writing 6502 assembly language on an Apple ][, and I loved it, but I hope I've forgotten enough of it to have room to learn new things. If only I could forget all those hex I/O and peripheral addresses from $C000-$CFFF and the Monitor ROM routines from $F800-$FFFF, without forgetting how brilliantly beautiful Woz and and Allen Baum's code is.
Forgetting old stuff to make room for new stuff is one of the most valuable skills you can have in this industry.
The next most important skill is persistence ;) -- writing stuff down before you forget it, in a way that won't make future-you hate present-you when you need to learn it again.
You forgot to read the last line of my message, in my profession I utilize LLMs to a high degree, my point was, that in my free time I code by hand 1) because self esteem benefits from the process of problem solving 2) I want to understand the fundamentals to not feel like a fraud 3) there is no time pressure to perform.
People neglecting this side will inevitably become less proficient at their job and become rather helpless without AI, that's all I'm saying.
Yeah, I read it, and we don't disagree about your job.
You put a number on it at the top of the thread: three years until people relying on agents have
forgotten how to do the job. Copilot's technical preview was 2021, ChatGPT was November 2022, and
agents that write the majority of the code are a 2024-2025 thing. So the window you're predicting
is roughly the entire span in which these tools have existed, and the atrophy hasn't shown up in it.
For scale, the last ARM I wrote was for a Sony Clié and the last PowerPC for a Mac, both more than
two decades ago. If I needed to write ARM for my Mac now, I could brush up about as fast as I could
on the 6502 I mentioned upthread, most of whose opcodes, $C000 I/O addresses and monitor ROM entry
points I genuinely cannot recall any more. Two to four decades of not touching an instruction set
didn't cost me the fundamentals, and it hasn't made me less proficient in Python and TypeScript or
helpless without AI. Three years of using a good tool is not going to do it either.
There's also a distinction buried in your last line, because "people neglecting this side will
inevitably become less proficient" bundles two different things. Losing a skill you had is not the
same as never acquiring one, and only the second is an argument about tools.
So take acquisition. I've been learning PDP-7 assembly in order to recover, OCR, and analyze 128
pages of PIXIE assembly source and octal machine code that Heinz Lemke wrote at Cambridge in the late 1960s,
and get it running in an emulator. A language I never knew, with custom instructions for bespoke
networking hardware, no Stack Overflow, and almost nobody left to ask except Heinz himself. AI has
been extremely helpful for exactly that, and the understanding is still mine to build: I have to
predict what the machine will do, and catch the explanations that are wrong.
On feeling like a fraud: the test isn't whether you typed it, it's whether you can predict what it
does and fix it when it breaks. Plenty of hand-typed code fails that test and plenty of carefully
reviewed generated code passes it. Your reasons 1 and 3 are good reasons to hand-code, since
problem solving is genuinely fun and nobody is waiting on your free time, but reason 2 is
measurable, so measure it instead of trusting the feeling.
The atrophy you're worried about is real. It just isn't caused by the tool. It happens to people
who stopped wanting to learn, and that predates AI by decades, even millennia. I watched it happen
back when much of the industry ran on assembly language.
If you're not using AI to teach yourself new things, you're not holding it right.
You're not wrong and it really comes down to how the tool is used. People wanting a quick solution, thus skipping the learning process entirely out of either complacency or because they feel like they don't have the time or energy are more at risk of atrophy I'd assume.
Personally the tools have simply gotten too good to ignore so I have been thinking about how to write skills and prompts in a way that enable me to keep learning, overcoming friction and challenges and not simply using the tool for hand holding and jumping straight to the end result.
For me personally, what matters is the process of learning and mastery, my self esteem depends on it, and enriching that with AI instead of AI eradicating that process for the sake of efficiency is my current challenge.
The thing is I think there will come a time when AI tools will not be as accessible as they are now. I just don't know if it is in 3 years or more, because the market can stay irrational longer than I can foresee.
I saw this idea from elsewhere that the things AI helps with was never the "profit bottleneck" of companies. 10x engineering productivity gain does not translate to 10x more revenue if your profit bottleneck is customer acquisition and retention. In other words, your profit is limited by how many people are willing to give you money for your services and AI can't really affect that.
And even the productivity gains per individual is a generous assumption. An individual can only prompt (and check! You guys check right?) so much. Sooner or later the pendulum is gonna swing in the other direction and it will be cheaper to build a team than equip individuals with AI _to deliver the same value_. I'm also assuming that AI is still in the VC-subsidized pricing stage.
Hence why I think the job in the future is gonna be pretty similar to the job four years ago.
> The next most important skill is persistence ;) -- writing stuff down before you forget it, in a way that won't make future-you hate present-you when you need to learn it again.
Huge amen. Another thing I've found Claude very useful for is writing documentation for legacy systems and cleaning up my own notes on it. I only need Claude to be 70-80% correct because from there I can take it. That error margin is no different from moderately-outdated-but-still-useful documentation.
When that pendulum swings, I will have a documentation binder that I will print money with.
> The thing is I think there will come a time when AI tools will not be as accessible as they are now
I am very convinced that today's SOTA level will be accessible 3 years from now, but you might be right about the SOTA accessibility at that point in time.
Still with today's SOTA you need to know much less details to be able to write code than without it.
My (arguably not very directly communicated, true) actual point is that "weird" is a term that is not really suited for tool choices like this.
"Weird" is a social concept. It's an artifact that has its roots in social cohesion and friction induced by individual nodes to glue the group together.
Software development otoh is an engineering discipline (or.. it should be).
And Engineering does not use the local social consensus algorithm for determining correctness. (or.. it should not)
I think they just meant its weird as in if you met a software dev at a dinner party who was still writing all their code by hand it would raise an eyebrow and you’d want to enquire further why that is.
I think understand the purpose of this comment. I think you're sensing friction and you'd like to reduce that by explaining for the other guy. Which is a common social script, but one that is being exploited here I'd say.
But, regardless, what is interesting I think is the scenario chosen there. Because a "dinner party" is not where engineering happens; and that's kinda the point I was getting at.
That's pulled from the pool of "social consensus" and not "math" or "physics" or whatever.
It's interesting, isn't it?
Just like how specific tokens in the context window of LLMs pull probabilities towards specific clusters of ideas (and tokens); with humans, you see similar things happen.
(This comment is more coherent than it might look at at first sight.)
___
Anyway, point (and to the point) is: Don't get hacked by the current thing fraudsters/grifters.
They don't respect you or me or really anyone the slightest. They're just in it to get rich quick (or simply just inflict psychological damage for sadistic reasons), and they will use any means necessary to do so.
This comment chain only exists because some hype guy is trying to induce FOMO into people for not doing more AI.
Why is unclear, but it's clearly malicious. Whether they consciously know that they are doing that doesn't matter for that assessment.
I think LLMs just allow you to get to where you want to get to quicker. So in a productivity oriented world it seems weird to not be making use of them where those gains can be had.
Let me try another analogy:
AI-related comment sections are like a watering hole for both get-rich-quick grifters, but also bad people wanting to hurt others.
The technology is so disruptive that there are all sorts of opportunities while stuff is still developing, things aren't regulated, and people themselves have not yet learned how to self-regulate.
So if you're looking to make money or to just inflict pain onto others, you're pulled towards this stuff.
This is what I am seeing in this specific comment sub-tree. A troll.
But in current year, you cannot just call people trolls, so instead you get these elaborate meta comments that bypass the anti-anti-troll defenses, but get so abstract, people just go "what".
My advice would be to view any of these comment sections through a zero-trust lens.
Godspeed. We will all eventually get through this.
Software can be an engineering discipline doesn't change whether it's a human doing it or AI doing it.
Arguably engineering is something AI should be really good at since you're just trying to compute a working product given constraints, formulas, resources.
Maybe "unusual" or "uncommon" would be better terms here.
I definitely wouldn't put it at 99% of developers, but I'd say it is the norm among developers I know that agents are their primary mode of writing code now (still using IDEs and GitHub to review changes).
The two people I know who don’t use coding agents work in government, and in a data science company working with government.
I’d say if you work at a tech company and agents aren’t writing the majority of your code, that is weird. But if you work at a traditional company that doesn’t have Claude or Codex subscriptions, there it might be pretty normal to still be writing code by hand.