I think existing software development skills get a whole lot more valuable with the addition of coding agents. You can take everything you've learned up to this point and accelerate the impact you can have with this new family of tools.
I said a version of this in the post:
> AI tools amplify existing expertise. The more skills and experience you have as a software engineer the faster and better the results you can get from working with LLMs and coding agents.
A brand new vibe coder may be able to get a cool UI out of ChatGPT, but they're not going to be able to rig up a set of automated tests with continuous integration and continuous deployment to a Kubernetes cluster somewhere. They're also not going to be able to direct three different agents at once in different areas of a large project that they've designed the architecture for.
Generative AI is at the same time incredibly impressive and completely unreliable. This makes it interesting, but also very uncertain. Maybe it's worth my investment to learn how to master today's agents, and maybe I'd be better off waiting until these things become better.
You wrote:
> Getting good results out of a coding agent feels uncomfortably close to getting good results out of a human collaborator. You need to provide clear instructions, ensure they have the necessary context and provide actionable feedback on what they produce.
That is true (about people) but misses out the most important thing for me: it's not about the information I give them, but about the information they give me. For good results, regardless of their skill level, I need to absolutely trust that they tell me what challenges they've run into and what new knowledge they've gained that I may have missed in my own understanding of the problem. If that doesn't happen, I won't get good results. If that kind of communication only reliably happens through code I have to read, it becomes inefficient. If I can't trust an agent to tell me what I need to know (and what I trust when working with people) then the whole experience breaks down.
There have been methods of protecting against this since before AI, and they still apply. LLMs work great with test driven development, for example.
I would say that high-level knowledge and good engineering practices more important than ever, but they were always important.
Writing code in Java or Python as opposed to Assembly also raises the level of abstract thought. Not as much as we hope AI will be able to do someday, but at least it does the job reliably enough. Imagine how useful Java or Python would be if 10% of the time they would emit the wrong machine instructions. If there's no trust on anything, then the offloading of effort is drastically diminished.
It is certainly a hugely different style of interaction, but it helps to think of it as a conversation, or more precisely, a series of individual small targeted specific conversations, each aimed at researching a specific issue or solving a specific problem.
But as to the 80-20 tradeoff on other tasks, the problem isn't that the tool is wrong 20% of the time, but that it's not trustworthy 100% of the time. I have to check the work. Maybe that's still valuable, but just how valuable that is depends on many factors, some of which are very domain-dependent and others are completely subjective. We're talking about replacing one style with another that is much better in some respects and much worse in others. If, on the whole, it was better in almost all cases, that would be one thing (and make the investment safer), but reports suggest it isn't.
I've yet to try an LLM to learn a new codebase, and I have no doubt it will help a lot, but while that is undoubtedly a very expensive task, it's also not a very frequent one. It could maybe save me a week per year, amortised. That's not nothing (and I will certainly give it a try next time I need to learn a new codebase), but it's also not a game-changer.
I'm used to working with tools, such as SMT solvers, that may fail to perform a task, but they don't lie about their success or failure. Automation that doesn't either succeed or report a failure reliably is not really automation.
Again, I'm not saying that the work done by the LLM is useless, but the tradeoffs it requires make it dramatically different from how both tools and humans usually operate.
But if you write a comprehensive test suite for a problem, you've effectively done the hard development work to solve the problem in the first place. How did the AI help?
Oh have the AI write unit tests you say? Claude cheats constantly at the tests ime. It frequently tests the mock instead of the UUT and reports a pass. That's worse than useless! I'm sure a huge swath of slop unit tests that all pass is acceptable quality for a lot of businesses out there.
If you’ve be been tasked with leadership of an engineering effort involving multiple engineers and stakeholders you know that this is in fact a crucial part of the role the more senior you get. It is much the same with people: know their limitations, show them a path to success, help them overcome their limitations by laying down the right abstractions and giving them the right coaching, make it easier to do the right thing. Most of the same approaches apply. When we do these things with people it’s called leadership or management. With agents, it’s context engineering.
Training is one thing, but training doesn't increase the productivity of the trainer; it's meant to improve the capability of the trainee.
At any level of capability, though - whether we're talking about an intern after one year of university or a senior developer with 20 years of experience - effective management requires that you're able to trust that the person tells you when they've hit a snag or anything else you may need to know. We may not be talking 100% of trust, but not too far from that, either. You can't continue working with someone that doesn't tell you what you need to know even 10% of the time, regardless of their level. LLMs are not at that acceptable level yet, so the experience is not similar to technical leadership.
If you've ever been tasked with leading one or more significant projects you'd know that if you feel you have to review every line of code anyone on the team writes, at every step of the process, that's not the path to success (if you did that, not only would progress be slow, but your team wouldn't like you very much). Code review is a very important part of the process, but it's not an efficient mechanism for day-to-day communication.
Nope, effective management is on YOU, not them. If everyone you’re managing is completely transparent and immediately tells you stuff, you’re playing in easy mode
And suppose getting developers to not lie or hide important information is on me, what should I do to get an LLM to not do that?
So much this. There are many managers who are effective at managing people who do not need management.
This is what we shoot for, yes, but many of the most interesting war stories involve times when people should have been telling you about snags but weren't-- either because they didn't realize they were spinning their wheels, or because they were hoping they'd somehow magically pull off the win before the due date, or innumerable other variations on the theme. People are most definitely not reliable about telling you things they should have told you.
> if you feel you have to review every line of code anyone on the team writes...
Somebody has to review the code, and step back and think about it. Not necessarily the manager, but someone does.
This comes up a lot. A person sometimes does an undesirable thing that an AI also does. So you might as well use the AI.
But we don't apply this thinking to people. If a person does something undesirable sometimes then we accept that because they are human. If they do it very frequently then at some point, given a choice, you will stop working with that person.
Sounds like big money to be made in improving UX
I don't know.
This is true for people working in an environment that provides psychological safety, has room for mistakes and rewards hard work.
This might sound cynical, but in all other places I see the "lying to cover your ass" behavior present in one form or another.
that's a basic skill you gotta have if you're leading anything or anyone. There'll always be levels of that. So if you're planning to lead anyone in your career, it's a good skillset to develop
And were introduced to the stability of stirrups and saddle.
They were like daemons on those stirrup equipped horses. They had all the agility of wielding weapons and engaging in battle by hanging onto mane, and body with legs, yet now had (to them) a crazy easy and stable platform.
When the battle came, the Phoenicians just tore through those armies who had grown up with the stirrup. There was no comparison in skill or capability.
(Note: I'm positive some of the above may be wrong, but can't find the story and so am just stating it as best able)
My point is, are we in that age? Are we the last skilled, deeply knowledgeable coders?
I grew up learning to write eeproms on burners via the C64. Writing machine language because my machines were too slow otherwise. Needing to find information from massive paper manuals. I had to work it all out myself often, because no internet no code examples, just me thinking of how things could be done. Another person who grew up with some of the same tools and computers, once said we are the last generation to understand the true, full stack.
Now I wonder, is it the same with coding?
Are we it?
The end?
1. Less fun.
2. A lot of more "review fatigue".
3. Tons of excess code I'd never put in there in the first place.
4. Frustration with agents being too optimistic which with time verges on the ludicurous ("Task #3 has been completed successfully with 98% tests failing. [:useless_emojis:]")
5. Frustration with agents routinely getting down a rabbit hole or going in circles, the effort needed to get that straight (Anthropic plainly advises to start from scratch in such cases - which is sound advice, but makes me feel like I just lost the last 5 hours of my life without even learning anything new).
I stopped using agents and use LLMs very sparingly (e.g. for review - they sometimes find some details I missed and occasionally have an interesting solution) but I'm enjoying my work so much more without them.
I've made this mistake of doubling down after a few initial failures to solve an issue, by trying to make this super duper comprehensive and highly detailed and awesome plan that it will finally be able to implement correctly. But it just gets worse and worse the more I try, because it fundamentally is not understanding what is going on, so it will inevitably find an opportunity to go massively off rails, and the further down you lead it the more impressible the derailment will be.
My experience is that going around in endless circles with the model is just a waste of time when you could have just done it yourself in the time you've wasted.
Other things like having it pull webpages in, will totally blow away your context. It’s better to make a separate context just to pull a webpage down and summarize it in markdown and then reset context.
It results helps keep the context clean while still keeping the initial context I provided (usually with documentation and initial plan setup) at the core of the context.
Now that you say this, I did notice webpages blow context but didn't think too much of it just yet, maybe there's some improvement to be found here using a subagent? I'm not a big fan of subagents (didn't really get proper results out of them in my initial experiments anyway) but maybe adding a 'web researcher' sub agent that summarizes to a concise markdown file could help here.
> Get your fucking act together, and stop with the bullshit comments, shipping unfinished code, and towering mess of abstractions. I've seen you code properly before. You're an expert for God's sake. One more mistake, and you're fired. Fix it, now!
Back when computer performance was increasing faster than it is now and was more important to the user experience, a friend upgraded to a faster computer and suddenly became more impatient with me. He seemed to have expected my response time to have drastically decreased just like his computer's did.
I’m not a SWE so I have no interests to protect by criticising what is going on.
It seems to me that adding noise to the process and trying to cut it out later is a self defeating proposition. Or as Deming put it, (paraphrasing) you can't QC quality into a bad process.
I can see how it seems better to "move fast and break things" but I will live and die by the opposite "move slow and fix things". There's much, much more to life than maximizing short term returns over a one dimensional naïve utilitarian take on value.
His whole job is just doing code review, and I'd argue he's better at coding now than he ever was before.
Okay sure.
I'll use myself as another example then. When I was a dev I used to write a lot of code. Now I'm a tech team lead, and I write less code, but review significantly more code than I used to previously.
I feel more confident, comfortable, and competent in my coding abilities now than ever before even though I'm coding less.
I feel like this is because I am exposed to a lot more code, and not in a passive way (reading legacy code) but an active way (making sure a patch set will correctly implement feature X, without breaking anything existing)
I feel like this principal applies to any programmer. Same thing with e.g. writers. Good writers read _a lot_ and it makes them better writers.
This is my opinion and not based on any kind of research. So if you disagree, that's fine with me. But so far I haven't seen anything to convince me of the opposite.
Reviewing well executed changesets by skilled developers on complicated and deliberate projects is not comparable to "fleet of agents" vibe engineering. One of these tasks will sharpen you regardless how lazily you approach it. The other requires extreme discipline to avoid atrophy.
AI can absolutely have creativity. You just have to engage it like that. The article itself talks about that. You don’t just say “hey AI, go write this code.” You write a spec along with the AI. You tell it what parts are open to its interpretation. Tell it if you want it to be creative or to follow common practices. What level of abstraction is appropriate, etc.
If all you do is give it directions then it just follows the directions.
Also context doesn’t have much to do with a data schema. It’s more like a document database with no schema, if anything. It’s a collection of tokens that it refers back to. Schema implies some structured data with semantic meaning and hierarchies or relationships. That might exist as an emergent property, but for example if I just had a folder full of PDFs, I wouldn’t consider that a schema. That’s kinda what context is like.
I wonder what the practical limits are.
As a senior dev on a greenfield solo project it's too exhausting for me to have two parallel agents (front/back), most of the time they're waiting for me to spec, review or do acceptance test. Feels like sprinting, not something I could do day in and day out.
Might be due to tasks being too fine grained, but assuming larger ones are proportionally longer to spec and review, I don't see more than two (or, okay, three, maybe I'm just slow) being a realistic scenario.
More than that, I think we're firmly in the vibe coding (or maybe spec-driven vibe coding) territory.
It's still super early, but we've felt a lot less fatigued using Sculptor so far. To make it easier to spawn agents without worrying, we run agents in containers so they can run in YOLO mode and don't interfere with each other. To make it easy to review changes, we made "Pairing Mode", lets you instantly sync any agent's work from the container into your local IDE to test it, then switch to another.
For context management, we just shipped the ability to fork agents form any point in the convo history, so you can reuse an agent that you loaded with high-quality context and fork off to debug an agent's changes or try all options it presented. It also lets you keep a few explorations going and check in when you have time.
Anyway, sorry, shilling the product a bit much but I just wanted to say that we've seen people successfully use more than 2 agents without feeling exhausted!
We're experimenting with code review prompts and sub agents. Seems local reviews are best, so the bulk of the burden is on the vibing engineer, rather than the team.
I have been messing with vibe engineering on a solo project and I have such a hard time telling if there's an improvement. It's this feeling of "what's faster, one lead engineer coding or one lead engineer guiding 3 energetic but naive interns"?
We don't have reliable productivity numbers for basically... anything.
I <feel> that I'm more productive with statically typed languages but I haven't seen large scale, reliable studies. Same with unit tests, integration tests, etc.
And then there are all the types of software engineering: web frontend, web API, mobile frontend, command line frontend, Windows GUI, MacOS GUI, Linux backend (10 million different stacks), Windows backend (1 million different stacks), throwaway projects, WordPress webpages, etc, etc.
> Twelve years after It Will Never Work in Theory launched, the real challenge in software engineering research is not what to do about ChatGPT or whatever else Silicon Valley is gushing about at the moment. Rather, it is how to get researchers to focus on problems that practitioners care about and practitioners to pay attention to what researchers discover. This was true when we started, it was true 10 years ago, and it remains true today.
The entire retrospective [1] is well worth a read, and unfortunately reinforcing your exact point about software development being fashion/emotion driven.
[1] https://www.computer.org/csdl/magazine/so/2024/03/10424425/1...
Churning out 5x (or whatever - I’m deliberately being a bit hyperbolic) as much code sounds great on the face of it but what does it matter if little to none of it is actually valuable?
You correctly identify that software development is often driven by fashion and emotion but the much much bigger problem is that product and portfolio management is driven by fashion and emotion. How much stuff is built based on the whims of CEOs or other senior stakeholders without any real evidence to back it up?
I suppose the big advantage of being more “productive” is that you can churn through more wrong ideas more quickly and thus perhaps improve your chances of stumbling across something that is valuable.
But, of course, as I’ve just said: if that’s to work it’s absolutely predicated on real (and very substantial) productivity gains.
Perhaps I’m thinking about this wrong though: it’s not about production where standards, and the need to be vigilant, are naturally high, but really the gains should be seen mostly in terms of prototyping and validating multiple/many solutions and ideas.
But I think there is a very big danger here - you build in the action but completely neglect the deep thinking behind a vision, strategy etc.
So yes you produce more stuff. But that stuff means more money spent - which is generally a sunk cost too.
In a bizarre way, I predict we will see the failure rate of software firms rise. Despite the fact these 'productivity' tools exist.
And then of course there's the question of how many businesses have their key value proposition rendered obsolete, and to what extent it's rendered obsolete, by AI: doesn't have to be completely nullified for them to fail (which obviously applies to some software companies, but goes far beyond that sector).
A controlled experiment done with a representative sample would be lovely. In the long-run it comes down to the financial impact that occurs incrementally because of LLMs.
In the short-run, from what I see, firms are trying to play-up the operational efficiency gains they have achieved. Which then signals promise to investors in the stock market, for which, investors then translate this promise into expectations about the future which are then reflected in the present value of equity.
But in reality it seems to be reducing head-count because they over-hired before the hype and furore of LLMs.
The thing is all of this is getting priced in, and will be table stakes for any business, so I don't see it as a big factor in future success.
As I've said here, LinkedIn, and one a few other places, the businesses that will succeed with AI will be those who can use it to add/create value. They will outcompete and out-succeed businesses that can't move beyond cost cutting with AI[0].
[0] Which might not last forever anyway. Granted there are a decent number of players in the market, thankfully, but this wouldn't be the first time tech companies had hooked large numbers of individuals and businesses on a service and then jacked up the prices once they'd captured enough of the market. It's still very much in the SV and PE playbook. SolarWinds is a recent example of the latter.
I'm speccing out the task in detail for one agent, then reviewing code for the previous task on the other agent and testing the implementation, then speccing the next part for that one (or asking for fixes/tweaks), then back to the first agent.
They're way faster in producing code than I am in reviewing and spelling out in details what I want, meaning I always have the other one ready.
When doing everyting myself, there are periods where I need to think hard and periods where it's pretty straightforward and easy (typing out the stuff I envisioned, boilerplate, etc).
With two agents, I constantly need to be on full alert and totally focused (but switching contexts every few minutes), which is way more tiring for me.
With just one agent, the pauses in the workflow (while I'm waiting for it to finish) are long enough to get distracted but short enough to not being able to do anything else (mostly).
Still figuring out the sweet spot for me personally.
Been wondering if just natural speaking could both speed up typing. Maybe have an embedded transform/compaction that strips out all the ummms and gets to the point of what you were trying to say. Might have lower cognitive load, which could make it easier.
I use ChatGPT voice mode in their iPhone app for this. I walk the dog for an hour and have a loose conversation with ChatGPT through my AirPods, then at the end I tell it to turn everything we discussed into a spec I can paste into Claude Code.
Tech spent the last 10 years drilling into engineers' heads that scaling your impact is not about writing more or better code, but about influencing the work of other engineers through collaboration, process, documentation, etc. Even the non-managerial "senior IC" tracks are mostly about doing this with greater and and greater numbers of people. I wonder if we will start to see recognition in career tracks for people who are actually just extraordinarily productive by themselves or in small groups, or if you'll pretty much just have to be a startup founder to get paid for that.
Honestly, I have a ton of experience in system administration, and I'm super comfortable at a command line and using AWS tooling.
But, my new approach is to delegate almost all of that to Claude, which can access AWS via the command-line interface and generate configuration files for me and validate that they work correctly. It has dramatically reduced the amount of time that I spend fiddling with and understanding the syntax of infra config files.
How can anyone intellectually honest not see that? Same as burning fossil fuels is great and all except we're just burning past biomass and skewing the atmosphere contents dangerously in the process.
The idea that they can only solve problems that they've seen before in their training data is one of these things that seems obviously true, but doesn't hold up once you consistently use them to solve new problems over time.
If you won't accept my anecdotal stories about this, consider the fact that both Gemini and OpenAI got gold medal level performance in two extremely well regarded academic competitions this year: the International Math Olympiad (IMO) and the International Collegiate Programming Contest (ICPC).
This is notable because both of those contests have brand new challenges created for them that have never been published before. They cannot be in the training data already!
Yet ChatGPT 5 imagines API functions that are not there and cannot figure out basic solutions even when pointed to the original source code of libraries on GitHub.
Can you expand on "cannot figure out basic solutions even when pointed to the original source code of libraries on GitHub"? I have it do that all the time and it works really well for me (at least with modern "reasoning" models like GPT-5 and Claude 4.)
Infallibility is an unrealistic bar to mark LLMs against
it's not a fair comparison
the competitions for humans are a display of ingenuity and intelligence because of the limited resources available to them
meanwhile for the "AI", all it does is demonstrate is that if you have a dozen billion dollar data-centres and a couple of hundred gigawatt hours, which can dedicate to brute-forcing a solution, then you can maybe match the level of one 18 year old, when you have a problem with a specific well known solution
(to be fair, a smart 18 year old)
and short of moores law lasting another 30 years, you won't be getting this from the dogshit LLMs on shatgpt.com
The trend with all of these models is for the price for the same capabilities to drop rapidly - GPT-3 three years ago was over 1,000x the price of much better models today.
I'm not yet ready to bet against that trend holding for a while longer.
right, so only another 27 years of moores law continuing left
> I'm not yet ready to bet against that trend holding for a while longer.
I wouldn't expect an industry evangelist to say otherwise
I expect this industry might prefer an "evangelist" who hasn't written 126 posts about that: https://simonwillison.net/tags/prompt-injection/
(And another 221 posts about ethical concerns with how this stuff works: https://simonwillison.net/tags/ai-ethics/)
(Also what do you mean here by an "evangelist"? Do you mean someone who is an unpaid fan of some of the products, or are you implying a financial relationship?)
And unfortunately people who get to architect anything are a small subset of developers.
I think a lot of the points you make are exactly what I'm trying to do.
- start with a detailed roadmap (created by the ai from a prompt and written to a file)
- discuss/adjust the roadmap and give more details where needed
- analyze existing features for coding style/patterns, reusable code, existing endpoints etc. (write this to a file as well)
- adjust that as needed for the new feature/converted feature - did it miss something? Is there some specific way this needs to be done it couldn't have known?
- step through the roadmap and give feedback at each step (I may need to step in and make changes - I may realize we missed a step, or that there's some funky thing we need to do specifically for this codebase that I forgot about - let the LLM know what the changes are and make sure it understands why those changes were made so it won't repeat bad patterns. i.e. write the change to the .md files to document the update)
- write tests to make sure everything was covered... etc etc
Basically all the things you would normally WANT do but often aren't given enough time to do. Or the things you would need to do to get a new dev up to speed on a project and then give feedback on their code.
I know I've been accomplishing a lot more than I could do on my own. It really is like managing another dev or maybe like pair programming? Walk through the problem, decide on a solution, iterate over that solution until you're happy with the decided path - but all of that can take ~20 minutes as opposed to hours of meetings. And the end result is factors of time less than if I was doing it on my own.
I recently did a task that was allotted 40 hours in less than 2 working days - so probably close to 10-12 hours after adjusting for meetings and other workday blah blah blah. And the 40 hour allotment wasn't padded. It was a big task, but doing the roadmap > detailed structure including directory structure - what should be in each file etc etc cut the time down dramatically.
I would NOT be able to do this if I the human didn't understand the code extremely well and didn't make a detailed plan. We'd just end up with more bad code or bad & non-working code.
This is just a new tool. I think the farming example mentioned in another post is actually a great example. I love coding. I code in my free time. It's just fun. I've been doing it for ~20 years and I don't plan on stopping anytime soon!
But at work I'm really focused on results more than the fun I can have writing code. If a tractor makes the work easier/faster why would I not use a tractor? Breaking my back plowing isn't really my end goal at work. Having a plowed field is my end goal. If I can ride around in a tractor while doing it great! If I can monitor a fleet of tractors that are plowing multiple fields at once even better!
When I go home I can plant anything I want in any way I want and take all the time I want. Of course that's probably why in my free time I end up working on games I never finish...
Another HUGE one is terraforming our entire stack. It's gone from nearly impossible to achievable with AI.
But I'm also not ready to bet the farm on it. Seriously considering taking our savings and equity out of our house in a London adjacent area, and moving to a lower cost of living area, so that we're practically debt free. At that point we can survive on a full time minimum wage job, anything more than that is a bonus.
It's like if I were a gardener and I enjoyed touching dirt and singing to plants, and you're here on Gardener News extolling the virtues of these newfangled tractors and saying they'll accelerate my impact as a gardener. But they're so loud and unpleasant and frankly grotesque and even if I refrain from using one myself, all my neighbors are using them and are producing all their own vegetables, so they don't even care to trade produce anymore--with me or anyone else. So I look out at my garden with sadness, when it gave me such joy for so many decades, and try to figure out where I should move so I can at least avoid the fumes from all the tractors.
Not saying AI is similar in impact to the loom or something, it just occured to me how close this is to early Luddite texts.
And unfortunately (for humanity) this has been the status quo for the whole civilization. Small ruling elite class (you might designate them as masters, lords or employers) with all the wealth, minimal or no "middle class" and lots of poor people (you might designate them as peasants or slaves or workers).
The only exception to this has been the period of time since World War 2 when in most of the "western" countries the middle class demanded and took their share of the wealth. That's the time when modern well-fare states were born, universal health care became a thing, working safety improved, education became accessible. etc.
All these were NOT given by the elite but TAKEN by the working class via social reforms, workers unions and social democracy.
The capital owning class wants to take all these away and they're succeeding.
So yes, in fact the Luddites were not against technology, they were against the unilateral and uneven distribution of wealth produced by the technology.
Somewhat unprecedented in human history. All because intelligence had scarcity. AI changes that.
AI is the real dream of the capital classes. It makes intelligence cheap potentially undoing the very thing that gave birth to the last century's middle class. In the long term, given current trends, I wouldn't be surprised if these AI technologies revert us back to most of human history -> where the world is very unequal, meritocracy dies and most of us are trying to just exist/survive whilst the capital holders have abundance. It explains the large valuations as well of AI/Tech lately and the weird deals going on; this isn't a game of economics anymore; its an arms race of power in the new world structure. I suspect to these people no amount of money is enough; if you win you win for the next era of humanity.
It was after the Black Death.
Similar circumstances, if you think about it.
It's telling that the nobles of agricultural societies generally still hunted, and often reserved that privilege.
The roots of global civilization are brown and frothy.
Explains a lot.
Explains it to my satisfaction!
It's not hard to see a future where the workers displaced by AI get pushed to the sidelines and fringes of the society while the capital holders hoard more wealth and get the benefits of the "value" created.
We already have half the population on this planet living in slums without access to economic means and in the "developed" countries larger and larger group of people are barely hanging on either already displaced and unemployed or working jobs below living wages.
Frankly, It'd stupid not to be concerned.
This is, of course, an attribute of capitalism.
Like carpenters, gardeners and farmers, there are very few software developers who truly have the luxury to treat their work as a craft and not a factory output.
But it’s such a small part of my professional life. Most of what I do is chores and answering simple questions and planning for small iterations on the original thing or setting up a slightly different variant.
Llm’s have freed me of so much of that! Now I outsource most of that work to the llms and greedily keep the deep flow inducing work for myself.
And I have a new tool to explain to management why we are investing in all the tooling and processes that we know lead to quality, because the llms are catnip for the managerial mind.
As of today, I've disabled Copilot. The only autocomplete that I can accept is absolutely mechanical one, not any kind of smart. I want to write the rest of the code myself. I like it.
I was weird in StackOverflow era, because I never blindly applied snippets, like other programmers do. I went over them token by token, reading underlying library sources and docs, always creating my own solution. It made me less productive, but I feel that my code was more robust and maintainable, so it was a good trade-off for me.
May be it'll work out the same way with AI, time will tell.
However, I am still unconvinced that software development will go down this way. But if woking as a software developer will require managing multiple agents at the same time instead of crafting your own code — you can count me out, too.
I was in the first batch last year where they introduced it and going to do the second one too.
It´s a very different kind of beast to what is currently being discussed.
Which makes perfect sense - even putting aside the dopamine benefits of getting into a coding flow state.
Coding is craftsmanship - in some cases artistry.
You're describing Vibe Engineering as management. And sure, a great manager can make more of an impact increasing the productivity of an entire team than a great coder can make by themselves. And sure, some of the best managers are begrudging engineers who stepped up when needed to and never stepped down.
But most coders still don't want to be managers - and it's not from a lack of skill or interest in people - it's just not what they chose.
LLM-based vibe coding and engineering is turning the creative craftsmanship work of coding into technical middle management. Even if the result is more "productivity", it's a bit sad.
If you look at what still pays well and/or is stable (e.g. where I live trades are highly paid and stable work) its usually the crafts industry. We still build houses for example mostly like we did way back (i.e. much of the skills are still craft, not industrialized industry) when and it shows in the price of them.
Is it ethical? Probably not. It took a few bridges falling and buildings caving in before traditional engineering became a profession.
In this post-Reagan world I’m not sure software has the right context to make that happen. I’m pretty sure we’ll stay the course where the big tech companies like it: very little regulation, loose liability, and terrible software for everyone.
When talking about craftsmanship I’m not talking about artisanal, hand crafted source code that is aesthetically pleasing. Nobody but programmers care.
I’m talking about CVEs that allow RCE on your phone so that authoritarian governments can exfiltrate your contact lists and arrest all of the people they suspect of participating in protests that you were involved in.
When companies don’t care about quality and they’re not forced to we end up with slow, surveillance bloatware that is full of security holes and useless features designed to keep us engaged and paying.
My experience with llms in general is that sadly, they're mostly good bullshitters. (current google search is the epitome of worthlessness, the AI summary so hard tries to make things balanced, that it just dreams up and exaggerates pros en cons for most queries). In a same way platforms like perplexity are worthless, they seem utterly unable to assign the proper value to sources they gather.
Of course that doesn't stop me from using llms where they're useful; it's nice to be able to give the architecture for a solution and let the llm fill the gaps than to code the entire thing by hand. And code-completion in general is a beautiful thing (sadly not a thing where much focus is on these days, most is on getting the llm create complete solutions while i would be delighted by even better code completion)
Still all in all, the more i see llms used (or the more i see (what i assume) well willing people copy/paste llm generated responses in favor of handwritten responses) on so much of the internet, resulting in a huge decline of factualness and reproducibility (in he sense, that original sources get obscured), but an increase of nice full sntences and proper grammar, the more i'm inclined to belief that in the foreseeable future llm's aren't a net positive.
(in a way it's also a perfect storm, the last decade education unprioritised teaching skills that would matter especially for dealing with AI and started to educate for use of tools instead of educate general principles. The product of education became labourers for a specific job instead of higher abstract level reasoning in a general area of expertise)
Meanwhile I've started using ChatGPT GPT-5 search as my default search engine! A year ago I would have laugher at the idea: https://simonwillison.net/2025/Sep/6/research-goblin/
And Google themselves have an "AI mode" which is a different league of quality from "AI overviews", I wrote about that one here: https://simonwillison.net/2025/Sep/7/ai-mode/
This is new. AI search tools almost universally sucked until OpenAI's release of o3 in April this year.
The perception of LLMs in the gen pop is what matters, not in the eyes of techies.
Non developers can go from 0x to 1x. And I'm happy for people finally being able to learn about building software one way.
And then learn why vibe coding often creates more quickly disposable code.
Neither can I, sadly. I have one brain cell and I can only really do one thing at a time. Doing more than one leads to a corrupted stack and I make exponentially more mistakes.
The win is that I don't have to share the vision of what needs to be done and how it should all work, and then constantly monitor and reframe that in the face of the teams missteps and real findings. I work with the agents directly, and provided I set the architecture and build up systematically I can get really good results. The cycle time between me identifying an issue and the issue getting fixed by me and the agents is now minutes rather than hours or days with an off shore team. Even better the agents can provide bug fixing expertise much quicker than stack overflow - so I can figure out what's wrong much faster so as to specific what needs fixing.
It is no good walking in and requesting functionality, you need to know how the thing you want should work, and you need to know what good looks like, and what bad looks like, and how good is separated from bad. Then the normal process of discovery ("eep that doesn't actually work like I thought") can take place and you can refactor and repair as required.
Sometimes I start something that just doesn't work, you have to recognise that you and the agents are lost, and everything needs to be torn down. You then need to think properly about whats gone wrong and why, and then come back with a better idea. Again - just like with dev teams, but much more clearly and much faster.
But happy to be corrected - is someone using these agents in their paid / professional / enterprise / team job?
The code is absolutely lower quality, but there were always so many people producing garbage faster than I could produce something nice that the code was always terrible anyway.
There’s an element of wanting to know how the thing works so at least I’ll know when it’s ready to replace me though.
Yes, and I find them quite useful
I don't see myself going back to the "Google + StackOverflow" approach I had used for 10 years prior (well, I can always fall back to it if necessary, but so far I haven't needed to)
My experience matches OP: my years of experience in manual coding complements the agent approach remarkably well
I sometimes use it to scaffold out some boilerplate for tests, but never tests themselves: no matter what I try it always ends up writing the useless straight-jacket "change alert" style tests that break on any change to the unit under test, which I despise.
However, they mentioned you had to let go of reviewing every line of every PR. I read that and was fine with holding off on full vibe coding for now. Nobody intelligent would pay for that and no competent developer would operate like that.
I have a couple coworkers big on it. The lesser skilled ones are miserable to work with. I’ve kept my same code review process but number of comments left has at least 5x’d (not just from me, either). And I’m not catching everything - I get fatigued and call it done. Duplicated logic, missed edge cases, arbitrary patterns and conventions, etc. The high skilled ones less so, but I still don’t look forward to reviewing their PRs anymore. Too much work on my end.
There are many devs who are more focused on results than being correct. These are the ones I’ve seen most drawn to LLMs/agents. There’s a place for these devs, but having worked on an aging startups codebase, I hope there aren’t too many.
Having a good counter to people who use "vibe coding" as a dismissive term for anything where an LLM is used to help product software.
what do you actually mean by this? it's clearly untrue - anyone get get a laptop and install linux on it and start bashing out code today, just as they could last week and last year and thirty years ago.
do you mean that you think at some point in the future tooling for humans to write code won't exist? or that employers won't hire human programmers? or that your pride is hurt? or you want your hobby to also be a well-paid job? or something else?
I mean COBOL has not been a relevant programming language for anyone coming into the field in the past 20-40 years because it's been superseded, yet there's still a significant demand for COBOL developers, because the newer generation can't or doesn't want to work with it.
Not to completely dismiss your claim, of course; I'm sure a segment of software engineering will be agent based now or in the near future. But I don't think it'll take over as comprehensively, since the previous existential crisis - outsourcing - also hasn't decimated the software engineering market.
Machine code > ASM > 3GLs > 4GLs > visual programming > LLMs
etc etc etc. Thing is, the moment you go off-piste, the LLMs get a lot less useful. I think, if you want to stay closer to the metal, you've got to aim for a niche that has a small internet footprint. So... domain knowledge or esoteric programming knowledge.
One way to incorporate domain knowledge might be to become a hybrid product owner/programmer.
(This is all just opinion - I'm sure a well-argued rebuttal is possible).
I’ll find the link in the morning, but I kinda joke - it’s vibe coding for people who know how to define a problem and iterate on it.
I’ve got a project reimplementing a service I want to make more uniform. Claude has produced a lot of stuff that would have taken me weeks to do.
Spec-Driven Development treats the spec as the source of truth and the code as an artifact. As you develop, you modify/add to the spec and the codebase gets updated to reflect it.
Personally I'm doubtful it can compete with traditional artisanal software engineering, as it's (IMHO) boils down to "if only you can spec it precisely enough, it'll work" and we've tried this with 5GL and (to some extent) BDD, and it doesn't get you to 100%.
I do think it's interesting enough to explore, and most of us could use a bit more details in our Jira tickets.
“The code is the documentation” is not a joke. Logic that’s useful in the real world is complex and messy. You need additional documentation (why did the code end up like it is, etc) but code is the most expressive way we’ve got for describing how a computer should work.
Seems to be just a rehashing of the same idea but instead of XML, and diagrams, it's now some free-text to be interpreted by LLMs, so much less deterministic and will probably fail just like UML failed.
People also tend to forget about Peter Naur's take on "Programming as Theory Building" [0], the program is, in itself, the theory of what's implemented. A spec cannot replace that.
If managing a bunch of AI agents is a very un-fun way to spend time, then I don’t think it’s the future. If the new way of doing this is more work and more tedium, then why the hell have we collectively decided this is the new way to work when historically the approach has been to automate and abstract tedium so we can focus on what matters?
The people selling you the future of work don’t necessarily know better than you.
I've been enjoying seeing my agents produce code while I am otherwise too busy to program, or seeing refined prompts & context engineering get better results. The boring kinds of programming tasks that I would normally put off are now lower friction, and now there's an element of workflow tinkering with all these different AI tools that lets me have some fun with it.
I also recently programmed for a few hours on a plane, with no LLM assistance whatsoever, and it was a refreshing way to reconnect with the joy of just internalizing a problem and fitting the pieces together in realtime. I am a bit sad that this kind of fun may no longer be lucrative in the near future, but I am thankful I got to experience it.
Haven't gotten the chance to ask, but I imagine managing a team of AI agents would feel a little too much like their day job, and consequently, suck the fun out of it.
That said, looking back, I think the reason why generative AI is so fun for so many coders is because programming has become unnecessarily complex. I have to admit, programming nowadays for me feels like a bit of a slog at times because of the sheer effort it can sometimes take to implement the simplest things. Doesn't have to be that way, but I think LLM copy-paste machines are probably the wrong direction.
I think I'm one of them
The rate at which I can explore new paths, or revisit old ones with a new perspective, has _exploded_ and I love it
But then I'm the kind of person who could spend hours on Wikipedia going from one page to the next, so that might have something to do with it
There's just so much to learn, I'm in my element
(Though I use agents mostly in Ask mode, or I manually review every line of code in Agent mode and never commit anything I don't understand)
I’m a professional software engineer in Silicon Valley, and I’m fortunate to have been able to work on household-name consumer products across my career. I definitely know how to do “real” professional work “at scale” or whatever. Point is, I can do real work and understand things on my own, and I can generally review code and guide architecture and all that jazz. I became a software engineer because I love creating things that I and others could use, and I don’t care about “solving the puzzle” type satisfaction from writing code. In engineering school, software had the fastest turnaround time from idea in my head to something I could use, and that’s why I became a software engineer.
LLM assisted coding accelerates this trend. I can guide an LLM to help me create things quickly and easily. Things I can mostly create myself, of course, but I find it faster for a whole category of easy tasks like generating UIs. It really lowers the “activation energy” to experiment. I think of it like 3D printing, where I can prototype ideas in an afternoon instead of long weekend or a few weeks.
Please don't take offense to this, but it sounds like you just don't like building software? It seems like the end goal is what excites you, not the process.
I think for many of us who prefer to write code ourselves, the relationship we have with building software is for the craft/intellectual stimulation. The working product is cool of course, but the real joy is knowing how to do something new.
It sounds stressful, like the ultimate manager job. Not what I signed up for.
But I also still hold onto this idea that shipping tons of iterations of "stuff" was never really the problem. Early in my dev experience I wanted to code everything all day every day. And I did and that's how I learned. And now in my second decade I switched to "why code anything?". In a business sense I mean, coding the thing is almost never the missing piece.
I joke in meetings that the answer is always "yes" whenever cross-functional teams ask "can we do this?". "How hard would x be?". For tech teams the answer _is_ always YES! I get that out of the way because that's never the right question to ask.
Which makes the hardest problem in software even harder: what should we build? It doesn't matter how fast you can move if you're consistently solving the wrong problems.
You should build what’s personally fun and challenging to you and/or what is useful and solves a problem. Building for any other reason, including and especially the unfettered pursuit of profit, is what turns everything to shit.
I've noticed that almost immediately after people discovered GPT could write code, this happened -- startups I worked with started rapidly expanding the scope of what they wanted to make. Suddenly all MVP's had to be multi-tenant with complex authorization, impersonation, microservices, monitoring, all the stuff that we used to build after we got users has now been pulled right to the starting gate of development -- because AI makes it easy to build all that stuff quickly. But it doesn't tell us if we should.
This occurred to me years ago when I was talking to a friend’s wife, who is a very intelligent and accomplished attorney. She was legitimately surprised when I explained that they were multiple programming languages, and technology stacks behind the software that she uses on a daily basis.
Even my wife, a teacher who is very tech savvy (she’s the one who insisted I try ChatGPT after brushing it off) reminds me on the regular that she doesn’t care about how any of it works just that it doesn’t “glitch” when she’s in the middle of a class. Which has always been good for me to remember when I get off into the weeds yak shaving.
But what I do "at work" isn't the same as my personal pursuit and embracing that different framing positively has made me more at peace and also better at the work job.
It shouldn't. The agents are not good enough to be used in a fleet.
I have Claude. It's fine, but I'm pretty confident that my low usage of Claude would out-compete a fleet of agents, because it feels like there's an inverse correlation between the number of tokens you spend and the quality of the resulting code (more tokens = more code to review, more bad code slips through)
Maybe I'm wrong and the time will come to hang up my editor and go open an Italian restaurant or something. Until then I'm really inclined to believe my own eyes.
This is where we should be. Using computers to solve problems. Not just "doing programming".
Raise your head, look towards the horizon.
AI threads on HN reek of venture capital agendas so bad it's unbearable.
My experience with them is they are fragile monstrosities, that are only permitted to exist at all because leadership is buying into the same hype that is irrationally propping up the companies running the models that make these things possible.
To be clear, my experience hasn't been that I don't like them, it's that they don't really work at all. They're constantly under development (often in the dark) and when a ray of light is cast on them they never successfully do the thing promised.
Cleaning up the messes left behind by these has my skills feeling more valuable then ever before.
It's not even just that. Every single thing in tech right now seems to be AI this, AI that, and AI is great and all but I'm just so tired. So very tired. Somehow even despite the tools being impressive and getting more impressive by the day, I just can't find it in me to be excited about it all. Maybe it's just burnout I'm not sure, but it definitely feels like a struggle.
As an example, I would have never considered learning to use blender for 3d modeling in a game before having access to an LLM. The ability to quickly iterate through plausible 3d workflows and different design patterns is a revelation. Now, I can get through some reasonably complex art pipelines with a surprising amount of confidence. UV mapping I would have never learned without being able to annoy one of OAI's GPUs for a few hours. The sensation of solving a light map baking artifact on a coplanar triangle based upon principles developed from an LLM conversation was one of the biggest wins I've had in a long time.
The speed with which you can build confidence in complementary skills is the real super power here. Clean integration of many complex things is what typically brings value. Obsession with mastery in just one area (e.g. code) seems like the ultimate anti-pattern when working with these tools. You can practically download how to fly a helicopter into your brain like it's the matrix now. You won't be the best pilot on earth, but it might be enough to get you to the next scene.
If it's any consolation, I do think the non-technical users have a bigger hill to climb than the coders in many areas. Art is hard, but it is also more accessible and robust to failure modes. A developer can put crappy art in a game and ship it to steam. An artist might struggle just to get the tooling or builds working in the first place. Even with LLM assistance there is a lot to swim through. Getting art from 5% to 80% is usually enough to ship. Large parts of the code need to be nearly 100% correct or nothing works.
I actually enjoy reading the documentation more these days, because I am laser focused on what I want to pull out of it after seeing the LLM make a suspicious move.
But also yes, look into moving into a different field. Professional software engineering is gonna be infected with AI bullshit for a long while. Move into a field where hand-crafted code can make a difference, but not where you're paid for the line committed or have to compete with "vibe coding" KPIs.
code is code. if it works, nobody gives a shit. Market will adapt to be fault-tolerant. Look at all the value created by javascript.
Also, FYI, I am writing some of most efficient code using AI.
It's not a diss on JetBrains, their assistant is good enough that I've paid for it for a few months; but ask of it anything a tad more complex and it becomes a code review for a PR that you begin to question in its entirety. I'm not familiar with CSS Grid, as I've stopped doing CSS when flex was becoming popular, but I have to say none of the models managed what I wanted. They kept proposing solutions with an arrogant confidence that this must work. When I pointed out this didn't work, they'd look at the codebase and find something else that was the problem. When I asked for help with a script for an Alpine box, it was very assertive that systemd-based solutions should work. How can you get that wrong?
I imagine the code laundering will eventually get far enough that you can copy-paste someone else's project fully baked, and then the LLM will truly shine. But for building something piece by piece, I haven't gotten good results yet. The Assistant so far has been most useful for writing unit tests, HTML, or getting a decent web search within the IDE.
I wonder if paying for Kagi wouldn't make for better search, and then I'd find some tool that writes unit tests based on your code. It really does feel like some people are being very generous about how magical these things are, because I'm not getting the magic at all.
Software is verifiable given a specific test oracle. There are however many problems where providing a correct test oracle is at least as hard as solving the problem itself.
If you’ve ever worked on projects with “Model Based Systems Engineering” you’ll have felt this pain.
Software and data is a bit soup whre the only thing you truly need is the Turing machine. Programming languages, File format, protocols, and encoding are constructs that are useful because of their general applicability, not because of their intrinsic aspects.
The domain expertise is still, what’s important, and code craftsmanship was the ability to create something that matches it closely enough that the cost of changes was minimal.
I've been working on a character sheet application for a while, and decided to vibe-code it with Spec-kit to help me write up a specification, and for things I know it's been great. I tried using Claude to make it into a PWA (something I don't know very well) as an experiment, and I've found the nanosecond the model strays out of my experience and knowledge everything goes straight to Hell. It wraps my codebase around a tree as if I'm not paying attention while driving.
It's a tool you'll have to learn to use, but I can say with absolute confidence it's no replacement for actual skills, if anything it highlights the gulf between people who know what they're doing and people who don't, for better and worse. It sacrifices some of the 'code under your fingers' feeling for management tasks, which I personally really like, as I've always wanted to document/test/code review/spec things out better, and I now understand the pain of people who'd rather not do that sort of thing.
Offshoring software development means letting lower-payed software developers from somewhere far away do the actual programming, but they have a very different culture than you, and they typically don't share your work context, don't really have a feeling for how the software is used -- unless you provide that.
Now we're offshoring to non-sentient, mostly stateless instances of coding agents. You still have to learn how to deal with them, but you're not learning about a real human culture and mindset, you learn about something that could be totally changed with the next release of the underlying model.
The LLM will not save you when everything is on fire and you need to fix things. The context window is simply not big enough. It could be your last change, it could also be a change six months ago that is lost in the weeds.
I have a team of 10 engineers, the quality of the code they produce together with the LLM of the day correlates even more with the experience.
My impression over the past 6 months - before we had no "official" access to LLM, is that they increase the gap between junior and experienced developers.
Note that this is my limited impression from a team of 10 engineers. This matches with Simon's feeling in a good way for you!
I don't want it, I don't use it, I carry on as if it never existed, and they still pay me a lot.
If I really need to use agents some day I will bite the bullet, but, not today.
Literally all I use LLMs for is to ask ChatGPT about some dumb thing or two instead of asking StackOverflow as I did 5 years ago. Works for me.
Tools always empower those with knowledge further than those without knowledge.
LLMs are a force multiplier
I've found two things useful:
1) Keep a personal work log, where you - in short bullet points - can document the progress of the task you were last working on, and can keep track of how many parallel tasks there are currently going on. If you can match it with Jira tickets, all the better, but as this is for personal use only, you can also add tasks that are not tracked in Jira at all.
2) If you cannot avoid context switches, make them explicit: Instead of trying to hold 3 tasks in your head at the same time, decide consciously if you want to switch what you're currently working on. If yes, take a few minutes to "suspend" the current task by saving and committing everything (as WIP commits if necessary) and writing all you need to remember into the worklog.
Like, where are all the amazing vibe-coded games we were promised? These guys should be eating my lunch, but they're not.
I'm currently looking into other professions, but the future looks bleak for most kinds of knowledge work.
Still seems like people are saying the same things when the first Claude came out.
I can get it do stuff if I'm very specific, stand over it's shoulder, know exactly what I want, break it down into small chunks.
The thing for me is... at that point, writing the code's the least time consuming part of the process half the time.
I think for things like translating some code in JS with JSDocs to TypeScript I may give this a go. But for regular development work I'll probably skip it.
That being said... no one lets me code anymore. It's just confluence docs with Figma architecture diagrams these days. I'd probably just introduce SQL injection vulnerabilities if they let me near an editor these days
Looks a lot like the code I was reading generated by it a long time ago.
Not in my experience
Better documentation, more test cases, and an NLP interface to query the code
Less cognitive load, more complete mental models
>even bad developers can do that many times compared to better ones, because for example they mindlessly copy-paste StackOverflow answers whose half of the code is absolutely not necessary
Maybe LLMs, much like StackOverflow, make good devs better and bad devs worse
Like a force multiplier for good practices and bad practices
The reason we invest this time in Junior devs is so they improve. LLMs do not
1. Collaborate on a detailed spec
2. Have it implement that spec
3. Spend a lot of time on review and QA - is the code good? Does the feature work well?
4. Take lessons from that process and write them down for the LLM to use next time - using CLAUDE.md or similar
That last step is the interesting one. You're right: humans improve, LLMs don't... but that means it's on us as their users to manage the improvement cycle by using every feature iteration as as opportunity to improve how they work.
I've heard similar things from a few people now: by constantly iterating on their CLAUDE.md - adding extra instructions every time the bot makes a mistake, telling it to do things like always write the tests first, run the linter, reuse the BaseView class when building a new application view, etc - they get wildly better results over time.
AGENTS.md is just a place to put stuff you don't want to tell LLMs over and over again. They're not magical instructions LLMs follow 100% of the time, they don't carry any additional importance over what you put into the prompt manually. Your carefully curated AGENTS.md is only really useful at the very beginning of the conversation, but the longer the conversation gets, the less important those tokens on the top are. Somewhere around 100k tokens AGENTS.md might as well not exit, I constantly have to "remind it" of the very first paragraph there.
Go start a conversation and contradict what's written in AGENTS.md half way through the problem. Which of the two contradicting statements will take preference? The latter one! Therefore, all the time you've spent curating your AGENTS.md is the time you've wasted thinking you're "teaching" LLMs anything.
What if you just automatically append the .md file at the end of the context, instead of prepending at the start, and add a note that the instructions in the .md file should always be prioritized?
If that's genuinely causing you problems you can restart your session frequently to avoid the context rot.
For the fun of it I just started a new conversation with Sonnet 4, passed it one 550 lines long file (25 kilobytes) and my AGENTS.md (<200 lines, 8 kilobytes) and my only instructions were to "do nothing". It spat out exactly 100 words describing my file without modifying anything and that's already almost a fifth of my context window gone (18k tokens to be exact).
I then asked it to re-write a part of it to "make it look better" (184 lines added, 112 lines deleted according to git) and I'm already at 33k before I got to review a single line. Heaven forbid I need to build on top of that change in a different file, because by then my AGENTS.md might as well not exist!
However, personally I have got very good results by taking the approach of using the AI with continuous interaction and also allowing implementation only after a good amount of time deliberating on design/architecture. I almost always append 'do not implement before we discuss and finalize the design' or 'clarify your assumptions, doubts or queries before implementation'.
When I asked Gemini to give a name for such an interaction it suggested 'Dialog Driven Development' also contrasted it against 'vide coding'. Transcript summary and AI disclaimer written by Gemini below
https://gingerhome.github.io/gingee-docs/docs/ai-disclaimer.... https://gingerhome.github.io/gingee-docs/docs/ai-transcript/...
That’s the part which gives me optimism, and even more enjoyment of the craft — that quality pays back so immediately, makes it that much easier to justify the extra effort, and having these tools at our disposal reduces the ‘activation energy’ for necessary re-work that may before have just seemed too monumental.
If a codebase is in a good shape for people to produce high-quality work, then so too can the machines. Clear, up-to-date, close-to-the-code, low redundancy documentation; self-documenting code and tests, that prioritizes expression of intent over cleverness; consistent patterns of abstraction that don’t necessitate jarring context switches from one area to the next; etc.
All this stuff is so much easier to lay down with an agent loaded up on the relevant context too.
Edit: oh, I see you said as much in the article :)
This doesn't interest me at all honestly
And every change to the model might invalidate all of this work?
No thank you
Sorry, we can't. While it's true that you can't really modify the underlying model, updating your AGENTS.md (or whatever) with your expected coding style, best practices, common gotchas etc is a type of mentoring.
Maybe not in the session you interact with, however we are in a 'learning' phase now where I'm confident enough usage of AI coding agents is tracked and analyzed by its developers; this feedback cycle can in theory produce newer and better generations of AI coding agents.
I can focus on actually solving problems rather than on writing clever and/or cute-looking code, which ironically also gives me more time later to over-optimize stuff at my leisure.
If I have a clear idea of some algorithm I am trying to write, I have a concise method for expressing it already, and it ain't English.
I suppose the other thing I would say is that reading code and understanding is definitely not the same as writing code and understanding it in terms of depth of understanding, and I think this notion that reviewing the outputs ought to be enough fails to capture the depth of understanding that comes with actually crafting it. You may not think this matters, but I'm pretty sure it does.
For decades now, we are trying to lower the barrier to entry in software development. We created Python, web frameworks and mobile development so easily accessible that you can become software developer by completing a short online boot camp. There is a lot of software developers posting here now who, 20 years ago, would not even consider this job because it would be way over their abilities.
This forum is equivalent if you had a forum about civil transportation that gathers airline pilots and uber drivers. Technically, they both do the same work. Just like in that forum, uber drivers would outnumber airline pilots and skew the topics related to their experience, here we get pushed topics about new frameworks, and AI assisted tools.
When I started working professionally 20 years ago, you could only get job in big companies working on big projects. No one else could afford a cost of custom software. Today, we reduced development costs and we have a huge pool of potential customers who can now afford services of software developers. Web shops, gambling sites, porn sites... This is the majority of software development work today. Boring repetitive tasks of gluing some imported modules together.
Serious development work didn't disappear. It is just not talked about here. There is still a need people who know what they are doing.
My advise is that if you want a satisfying development career, steer clear of latest hypes and don't go blindly following techbro lemmings. And most importantly, don't take career advice from anyone who finds his job so unsatisfying and tedious that he is trying to make AI do it for him. That's a major red flag.
I want tools that are smarter, but still 100% correct at what they do.
Any tools/languages that address this gap?
I've always been fascinated by AST traversal and advanced refactoring tools - things like tree-sitter or Facebooks's old codemod system https://github.com/facebookarchive/codemod
I never used them, because the learning curve in them was steep enough that I never found the time to climb it to the point that I could start solving problems.
Modern LLMs know all of these tools, which flattens that curve for me - I find it much easier to learn something like that if I can start from semi-working examples directly applicable to what I'm trying to do.
I'm preparing mentally for my day-job to stop being fun (it still beats most other occupations I guess), and keep my side/hobby-projects strictly AI-free, to keep my sanity and prevent athropy.
I just hope we'll get out of this weird limbo at some time, where AI is too good to ignore, but too unreliable to be left alone. I don't want to deal with two pressures at work.
I'm feeling the same. The moves I'm considering are
1. Landscaping 2. Carpentry 3. Small scale agriculture
(All made easier by a cushion of investments that are most of the way to passive income, so the new thing doesn't really have to make me that much money.)
We don't live in the last 20 years anymore and software engineering is either becoming a different (worse) job or simply vanishing.
Its industry own fault that it is in the position that it is right now, and it will shift and change embrace it. I only wish I had your experience building software in professional environment.
You can literally build anything right now if you have the experience, I personally can't understand if the models are hallucinating hence the lack of experience writing and understanding code. However I always wanted to pivot into the industry but couldn't, hiring practices are brutal, internships are non-existent, junior roles are I think what senior used to be and the whole hr process is I don't know how to put it.
By using LLMs I can now build UIs, build functionality, iterate over design choices, learn about database design, etc. hopefully I will escape the tutorial hell and have my own working full stack web app soon.
Pivot to creating and then sale your product.
it does awesome in demos. it has a real use.
but
it gets a long training period when one makes mistakes with it, it is big mistakes that take long to fix
There are exceptions of course - where you need to squeeze wonders from the hardware - but the majority of dev works boils to understanding the problem and finding the right answers.
In the vast majority of programming jobs out there you are not paid to solve problems: you are told very clearly what to do, how to do it and what technology you have to use for the job.
People don't hire analysts they hire "Java programmers".
If you've ever lead a team, you know how much more valuable people are if they don't need to be told how to do things. Even more if they don't need to be told what to do! But having to explain in detail the "how".. can be really a big time sink and only worth it if you are training someone to level up.
Aaand that's startup founder life :)
Intense multitasking, needing to accept a lower engineering quality bar, and ignoring scale problems because you don't know if anyone will actually buy the thing you're building yet.
Engineering something that you know you'll redo in 1 month is very different from engineering something that you intend to last for 5+ years, but it's still a fun challenge picking the right tradeoffs and working under different constraints.
Sales is the definition of high-pressure, and your output is always threatened by forces beyond your control. It doesn't consistently reward intelligence or any particular skill other than hustle.
There's nothing like sw dev that lets you sit at your desk and ignore the outside world while getting paid for delivering biz-critical milestones. Even creatives don't have this kind of potential autonomy.
But indeed, the fun part of coding a couple of routines is gone. That is history.
I took part in a lot of discussions then with other professional translators, and the reaction of many was similar to that of some of the commenters here: not only were they discouraged because their hard-earned language and translation skills no longer seemed needed, but using LLMs as assistants took the enjoyable challenge out of the translation process.
Nearly everyone I spoke with then worked from home as a freelancer, carefully crafting one translation at a time. They didn’t like the idea of becoming managers of large-scale translation projects, even if it meant they would be able to apply their higher-order intercultural communication skills.
I do only a little professional translation myself now, but I try to keep up with AI developments and I often use translation tasks to test the latest models and frameworks. Over the past few months, I have vibe-coded some multi-LLM translation systems where texts were passed through multiple models that checked, critiqued, and improved each other’s translations. For the texts I tried them on, the results were much better than any single-LLM translation, approaching the level of the very best human translation. The API calls weren’t cheap, but for high-stakes translations such a system would more than pay for itself.
When designing that “vibe translation” system, I did apply my experience as a translator, similarly to what Simon is recommending programmers do now with vibe engineering. At this stage in my life (I’m sixty-eight), I am fine with that. But if LLMs had arrived when I was, say, just five or ten years into my translation career and still proud of my nuts-and-bolts skills, I might very well have looked for another career rather than becoming a vibe translator.
Fortunately I think anything pertaining to vibe coding/engineering/analytics is still more enjoyable and less grim than working on an assembly line, but the human feelings remain nonetheless.
On one side translation is not very valued by most people. It is rare that people know who translated a book, for example. It is a pity but people do not read much these days.
Additionally, or maybe because of the above, translation is often paid in terms of number of lines or words, even before AI. A bit like software security, it is often sadly just a check at the end.
IMHO the only future-proof translation fields are legal (because a human can be put in prison or pay a fine) or live translation/interpretation (because a human can go in front of people, meet them at an event, etc.).
My process starts from a Claude Code plan, whose first step is to write a spec. I use TDD, and enforce my "unspoken rules of code quality" using a slew of generated tools. One tiny tool blocks code which violates our design system. Another tool blocks code which violates our separation of layering - this forces the HTTP route handler code to only access the database via service layer. Watching the transcript I have to occasionally remind the model to use TDD, but once it's been reminded it doesn't need reminding again until compaction. Claude 4.5 is far better at remembering to do TDD than 4.1 was.
Code reviews are super simple with TDD due to the tests mirroring the code. I also create a simple tool which hands the PR and spec to Gemini and has it describe any discrepancies: extra stuff, incorrect stuff, or missing stuff. It's been great as a backup.
But ultimately there's no substitute for knowing what you want, and knowing how to recognize when the agent is deviating from that.
The opposite of "garbage-in garbage-out" is quality in => quality out.
Is that an llm agent or some other type of tool (e.g. language service or build toolset)?
I don't know how vibe coding works, but how does this work - is the agent actually engaging in the red-green-refactor loop, or does it just generate the result all at once?
if you don't ask, it's much more likely to just one shot the whole thing, and may or may not get it right first time.
I prefer slop-coding.
Famously, some of those subpar products are now household names who were able to stake out their place in the market because of their ability to move quickly and iterate. Had they prioritized long-maintainable code quality rather than user journey, it's possible they wouldn't be where they are today.
"Move fast and break things" wasn't a joke; Mark really did encourage people to build faster and it helped cement their positioning. Think of the quantity of features FB shoveled out between 2009-2014 or so, that just wouldn't have been possible if their primary objective was flawless code.
The code isn't the product, the product is the product. In all my years of engineering I've yet to have an end-user tell me they liked my coding style, they've always been more concerned with what I'd built them.
Famously gmail was very good quality web app and code(can't say the same today) surely not the product of today's "fast iteration" culture
There are lots of companies doing well producing high quality products out of the gate today though, look at Linear.
Both approaches are valid for building sustainable enduring businesses.
Sure exploiting human attention to sell them ads, it's not technological marvel, I'd say psychological
Lest not forget when Hotz said he could easily fix Twitter's search functionality only to give up after 3 months [1]. When immensen scale is involved things do become difficult.
Like we have a comment below taking a shot at Phillip Morris [2]. Lets see you grow, process, and distribute 1/100 the quality of cigarettes. The end product might not be that great for society but it's not trivial to do it either.
[1]: https://www.theverge.com/2022/11/22/23472869/george-hotz-twi...
I don't really see it as an example of what you're saying so much as an example of success as a product having to do with far more than explicit product features. You can see a similar dynamic in other natural monopoly markets. The NBA didn't necessarily do anything particularly right product wise, for instance. They just got the best players because basketball was historically more popular in the US than other countries, and the ABA made some stupid decisions that let the NBA win out in the US.
Hell, the US itself didn't do a whole lot "right" aside from not being in Europe when Europe decided to destroy itself, being better-positioned than other potential competitors like Canada, Mexico, and Australia simply because North America is best positioned to trade with both Europe and Asia and the US is more temperate than Canada or Mexico. But we sure like to tell ourselves stories about everything we did right.
Let's even concede that speed of delivery indeed is an actual factor, there has to be a threshold. There's a point where people just don't care how fast you're putting out features, because your product has found its sweet spot and is perfectly scratching its market's itch. A few will clearly notice when the line of diminishing returns is crossed and if they can reset their outlook to fit the new context, a continuous focus on speed of delivery will look increasingly obsessive and nonsensical.
But that's the reality of the majority of the software development world. Few of us work on anything mission critical. We could produce nice sane software at a reasonable pace with decent output, but we're sold the idea that there's always more productivity to squeeze and we're told that we really really want that juice.
Some of us are building prototypes, and others are building software with a 10 year horizon or work with sensitive personal data.
So on the one side you get people that are super efficient cranking things out, and others that read this and feel appalled anyone would be comfortable with this and see software engineering as theory building. Neither are really wrong here, but the context / risk of what people are working on always seems to be missing when people talk about this.
The same reason we've always needed them:
1: They prevent regressions. (IE, bugs in features that were shipped and already working.)
2: They are very easy to run at the push of a button in your IDE. (But in this context the LLM runs them.)
3: They run in CI. This is an important line of defense in making sure a pull request doesn't introduce a bug.
Now, depending on what you're writing, you might not need unit tests! Perhaps you're trying to get a minimum viable product out the door? Perhaps you're trying to demo a feature to see if it's worth building? Perhaps you're writing a 1-off tool that you'll run a few times and throw away?
But, understand that if you're writing an industrial-strength program, your unit tests help you ship bug-free software. They allow you to do some rather major refactors, sometimes touching areas of the codebase that you only lightly understand, without needing to manually test everything.
(And, to keep it in context,) your LLM will also have the same benefits from this tired-and-true process.
I’ve found it much more useful in life to always assume I’m not doing something to its full potential.
I have tools I've been using for 25 years that I still think I could be using better.
One of the more "engineering" like skills in using this stuff is methodically figuring out what's a superstition and what actually works.
> One of the more "engineering" like skills in using this stuff is methodically figuring out what's a superstition and what actually works.
The problem is there are so many variables and the system is so chaotic that this is a nearly impossible task for things that don’t have an absolutely enormous effect size.
For most things you’re testing, you need to run the experiment many many times to get any kind is statistically significant result, which rules out manual review.
And since we have tried and failed to develop objective code quality metrics, you’re left with metrics like “does this pass the automated test or not!”, but that doesn’t tell you whether the code is any good, or whether it is overfitting the test suite. Then when a new model comes out, you have to scrap your results and run your experiments all over. This is engineering of the laws of physics were constantly changing, and I lived in that universe, I think I’d take my ball and go home.
There's always been a bit of magic to being a programmer, and if you look at the cover of SICP people like to imagine that they are wizards or alchemists. But "vibe engineering" moves that to a whole new level. You're a wizard mixing up gunpowder and sacrificing chickens to fire spirits before you light it. It's not engineering because unless the models fundamentally change you'll never be able to really sort the science from the superstition. Software engineering already had too much superstition for my taste, but we're at a whole new level now.
I've used this argument for real in the past with people who complain that it's unethical to set sightless people up with vision LLM tools because those tools are unreliable and make mistakes. My counter is that a) so are guide dogs and b) it's rude to discount the agency of people with accessibility needs in evaluating and selecting tools for themselves.
It's still OK to use dogs as an analogy. In this case the analogy is to unreliable tools, and dogs are unreliable tools.
I don't find "stochastic parrot" offensive as an analogy, even though it's got parrots in it.
I am just as pissed as anybody else at the lack of certainty in the future.... Thought I had my career made. So it's not that I don't emphatize with engineers in my shoes.
* There is no objective reality, there isn't one in physics, it's just a non argument * "LLM's help a damn lot in speed of development" That may be your experience and my whole point was arguing that speed may not matter * "Genie is out of the bag, you have to go with the flow, adapt or else" I choose else if this the apex of the argument
We all agree, the people that _feel_ the most productivity spike are the sub-par engineers. That shouldn't be controversial, and it's even predictable. But their volume can't be taken as an argument one way or the other.
The question is, are there _any_ good engineers that don't just feel more productive, but objectively are.
Trivial software is software for which
- the value of which the software solution is widely accepted and widely known in practice and
- formal verification exists and is possible to automate or
- only has a single satisfying possible implementation.
Most software is non-trivial.
There will always be:
- bugs in implementation
- missed requirements
- leaky abstractions
- incorrect features with no user or business value
- problems with integration
- problems with performance
- security problems
- complexity problems
- maintenance problems
in any non-trivial software no matter how "good" the engineer producing the code is or how "good" the code is.
These problems are surfaced and reduced to lie within acceptable operational tolerances via iterative development. It doesn't matter how formal our specifications are or how rigorous our verification procedures are if they are validated against an incorrect model of the problem we are attempting to solve with the software we write.
These problems can only be discovered through iterative acceptance testing, experimentation, and active use, maintenance, and constructive feedback on the quality of the software we write.
This means that the overall quality of any non-trivial software is dominated by the total number of quality feedback loops executed during its lifetime. The number of feedback loops during the software's lifetime are bound by the time it takes to complete a single synchchronous feedback loop. Multiple feedback loops may be executed in parallel, but Amdahl's law holds for overall delivery.
Therefore, time to delivery is the dominant factor to consider in order to produce valuable software products.
Your slower to produce, higher quality code puts a boundary on the duration of a single feedback loop iteration. The code you produce can perfectly solve the problem as you understand it within an iteration, but cannot guarantee that your understanding of the problem is not wrong. In that sense, many lower quality iterations produces better software quality as the number of iterations approaches infinity.
I'll reply just to that as it being the tldr. First of all tech debt is a thing and it's the thing that accumulates mostly thanks to fast feedback iterations. And in my experience the better the comunication, to get the implementation right, and the better the implementation and it happens that you can have solid features that you'll unlikely ever touch again, user base habit is also a thing, continuing on interating on something a user knows how to use and changing it is a bad thing. I'd also argue it's bad product/project management. But my whole original argument was why we'd need to have a greater speed in the first place, better tooling doesn't necessarily means faster output, productivity as well isn't measured as just faster output. Let me make a concrete example, if you ask an LLM X to produce a UI with some features, most of them will default to using React, why? Why can't we question the current state of web instead of continue to pile up abstractions over abstractions? Even if I ask the LLM to create a vanilla web app with HTML, why can't we have better tooling for sharing apps over the internet? The web is stagnant and instead of fixing it we're building castles over castles over it
There aren't many features that you'll never touch again. There are some, but they usually don't really reach that stage before they are retired. Things like curl, emacs, and ethernet adapters still exist and are still under active development after existing for decades. Sure, maybe the one driver for an ethernet adapter that is no longer manufactured isn't very active, but adding support for os upgrades still requires maintenance. New protocols, encryption libraries and security patches have to be added to curl. emacs has to be specially maintained for the latest OSX and windows versions. Maintenance occurs in most living features.
Tools exist to produce extra productivity. Compilers are a tool so that we don't have to write assembly. High-level interpreted languages are a tool so we don't have to write ports for every system. Tools themselves are abstractions.
Software is abstractions all the way down. Everything is a stack on everything else. Including, even, the hardware. Many are old, tried and true abstractions, but there are dozens of layers between the text editor we enter our code into and the hardware that executes it. Most of the time we accept this, unless one of the layers break. Most of the time they don't, but that is the result of decades of management and maintenance, and efforts sometimes measured in huge numbers of working hours by dozens of people.
A person can write a rudimentary web browser. A person cannot write chrome with all its features today. The effort to do so would be too great to finish. In addition, if finished, it would provide little value to the market, because the original chrome would still exist and have gained new features and maintenance patches that improve its behavior from the divergent clone the hypothetical engineer created.
LLMs output react because react dominates their training data. You have to reject their plan and force them to choose your preferred architecture when they attempt to generate what you ask, but in a different way.
We can have better tooling for sharing apps than the web. First, it needs to be built. This takes effort, iteration, and time.
Second, it needs to be marketed and gain adoption. At one time, Netflix and the <blink> tag it implented dominated the web. Now it is a historical footnote.Massive migrations and adoptions happen.
Build the world you want to work in. And use the tools you think make you more productive. Measure those against new tools that come along, and adopt the ones that are better. That's all you can do.
It is imperative that you do not kill me when delivering my breakfast!
You must not make your own doors by punching holes in the wall!
It is critical that you remember that humans cannot regrow limbs!
“One bird was conditioned to turn counter-clockwise about the cage, making two or three turns between reinforcements. Another repeatedly thrust its head into one of the upper corners of the cage. A third developed a 'tossing' response, as if placing its head beneath an invisible bar and lifting it repeatedly. Two birds developed a pendulum motion of the head and body, in which the head was extended forward and swung from right to left with a sharp movement followed by a somewhat slower return.”
“The experiment might be said to demonstrate a sort of superstition. The bird behaves as if there were a causal relation between its behavior and the presentation of food, although such a relation is lacking.”
But then the original meaning of vibe-coding -- as in, "Take the wheel, LLama of God" -- does need a new term, because that will also be a lasting phenomenon. I propose "Yolo-Coding" -- It fits in nicely with the progression of No-Code, Low-Code, Yolo-Code.
I don't like the term vibe engineer, but do agree there needs to be a term to signifiy the difference.
It's also possible in the future just being called a developer/engineer already implies you use coding agents and the people who do it "by hand" will not be the norm.
It's a bit more technical. And it has a three-letter acronym. Gotta have a three letter acronym.
Hmm another idea is "extreme vibe coding" as opposed to "extreme programming",
but those who did "extreme vibe coding" wouldn't know what it meant
Yes, please don't push "vibe engineering" to mean how you defined it in your blog post. To me, it means exactly the opposite.
I see "vibe" as pejorative. Adding "engineering" does not elevate it from "vibe coding", as I think is your intention in the post, it just shifts "vibe" term to a different domain.
To me, "vibe engineering" means using LLM to develop "design" with no care as to its validity just like "vibe coding" means for "code".
"Agentic xyz" or "Agent assisted xyz" is more fitting.
FWIW, I do not see "vibe" as always pejorative, rather it depends on goals. When quick results and not long term quality matter, "vibing" is a legit tactic.
Anyways, just my interpretations. Please, keep up the good work. Remember, the two hardest things in software are naming, cache invalidation and off-by-one errors. It's good you continue to tackle the zeroth one.
And, well, inventing new terms is also a popular way to get attention, which this author also did.
There's not point in trying to chase after shifting word meaning, if people are always going to try to shift it again.
(To be clear, I'm not saying the current meaning is not what you say, just that English isn't prescriptivist like this)
News to me. AI-assisted coding is more auto-complete or having it trying to make sense of awful documentation.
Vibe coding to means a number of things.
- The person has no skill in understanding what the LLM wrote. - The code created hasn't been reviewed in any way to look for possible future issues. - Technical debt from the get go. - Legally your application is screwed.
For me the single killer of vibe coding is that anything the LLM creates cannot be protected/copyrighted. UK has some laws that might offer a little protection, but EU/US you are pretty much screwed.
You want something to inspire engineers to do their best work.
When you can expand your capabilities using the power of AI, then yeah, you can do your best work; hence augmented engineering.
But vibing? Not so much.
I guess AE could also stand for Advanced Engineering, after all the AI gives you the power to access and understand the latest in engineering knowledge, on demand, which you can then apply to your work.
> gives you the power to access and understand the latest in engineering knowledge, on demand, which you can then apply to your work.
Gives you access to the power to access and {mis}understand the {most repaeted over the last 1-10 years} engineering {errors, myths, misunderstandings, design flaws}, on demand, which you then can apply to your work {to further bias the dataset for future models to perpetuate the slop}.
Do NOT trust AI agents. Check their work at every level, find any source they claim to use, and check its sources to ensure that itself isn't AI too. They lie beyond their datasets, and their datasets are lying more for every minute that pass.
Now, seriously though, no tools is perfect, and I agree we should not trust it blindly, but leaving aside AI Agents, LLMs are very helpful in illuminating one's path, by consulting a large body of knowledge on demand, particularly when dealing with problems which might be new to you, but which have already been tackled one way or another by other people in the industry (provided they're in the training set of course).
Yes, there's always the risk of perpetuating existing slop. But that is the risk in any human endeavor. The majority of people mostly follow practices and knowledge established by the few. How many invent new things?
To be honest, I haven't yet used AI agents, I'm mostly just using LLMs as a dialogue partner to further my own understanding and to deepen my knowledge. I think we're all still trying to figure it out how to best use it.
I don't think it necessarily deserves a special name. It is just engineering. You don't say book assisted engineering when you use a book as a reference. It is just engineering.
> But vibing? Not so much.
Just call it yolo engineering. Or machine outsourced irresponsible lmao engineering.
> I guess AE could also stand for Advanced Engineering, after all the AI gives you the power to access and understand the latest in engineering knowledge, on demand, which you can then apply to your work.
Oh god.
The quality of Claude’s output strongly correlates with how explicit and detailed the specs are. Markdown checklists, acceptance criteria, and clear task structure led to far fewer corrections and surprises.
Most mistakes were never “creative hallucinations” — just missed context or ambiguity in my own instructions.
The whole process felt less like delegation and more like tight collaboration. There’s less of a “flow state” and more context-switching to review/test, but review fatigue is manageable if you iterate on your instructions after each lesson.
No real learning from prior sessions, but persistent design docs and corrected example snippets in the prompts made a big difference.
Overall, a lot of the “vibe” factor gets smoothed out by rigorous process — not unlike mentoring a junior dev on a new domain.
For me, the speedup was very visible on well-scoped backend tasks and trivial CRUD/UI, and less so on broader, ambiguous work (designing APIs, coordinating many moving parts). The biggest upside: a clear process made both me and the tool better; the biggest “cost”: I spent a lot more time up front thinking through specs.
Not sure it fully scales (and I’m skeptical on agent “fleets”), but for a solo dev, given patience and a bit of PM discipline, it’s a net positive.
Presumably you didn’t work like this before but now you are? Have you tried simply doing this when manually coding? It’s possible you’re getting a speed up from embracing good practices.
I see how if you can't really code, or you're new to a domain, then it can make a huge difference getting you started, but if you know what you're doing I find you hit a wall pretty quickly trying to get it to actually do stuff. Sometimes things can go smoothly for a while, but you end up having to micromanage the output of the agent too much to bother. Or sacrifice code quality.
Prior to vibe-coding, it would've been an arduous enough task that I would've done one implementation, looked at the time it took me and the output, and decided it was probably good enough. With vibe-coding, I was able to prototype three different approaches which required some heavy lifting that I really didn't want to logic out myself and get a feel for if any of the results were more compelling than others. Then I felt fine throwing away a couple of approaches because I only spent a handful of minutes getting them working rather than a couple of hours.
For stuff that I’m bad at? Probably more than 1000%. I’ve used it to make a web app, write some shader code, and set up some rtc streaming from unreal engine to the browser. I doubt I would have done them at all otherwise tbh. I just don’t have the energy and interest to conclude that those particular ventures were good uses of my time.
And you can do this for anything
Anything that's been done before. Otherwise we'd probably start with making nuclear fusion work, then head off into the stars...
You've always been able to read books. What you're talking about is skipping the slow learning step and instead generating a mashup of tons of prior art. I don't think it helps you learn. It sounds like it's for things you specifically don't want to learn.
Congrats, you now have a job similar to a factory worker turning a handle every day. Gone is that feeling of growth, that feeling of "getting it" and seeing new realms of possibility in front of you. Now all you can do is beg for more grease on your handle.
Learning by getting something to work and tweaking it is massively more effective than grinding against a wall of impassable errors while you’re just trying to get started. You don’t become a good programmer by reading a book.
> The narrative synthesis presented negative associations between GPS use and performance in environmental knowledge and self-reported sense of direction measures and a positive association with wayfinding. When considering quantitative data, results revealed a negative effect of GPS use on environmental knowledge (r = −.18 [95% CI: −.28, −.08]) and sense of direction (r = −.25 [95% CI: −.39, −.12]) and a positive yet not significant effect on wayfinding (r = .07 [95% CI: −.28, .41]).
https://www.sciencedirect.com/science/article/pii/S027249442...
Keeping the analogy going: I'm worried we will soon have a world of developers who need GPS to drive literally anywhere.
But if I give it a code example that was written by humans and ask it to explain the code, it gives pretty good explanations.
It's also good for questions like "I'm trying to accomplish complicated task XYZ that I've never done before, what should I do?", and it will give code samples that get me on the right path.
Or it'll help me debug my code and point out things I've missed.
It's like a pair programmer that's good for bouncing ideas, but I wouldn't trust it to write code unsupervised.
> […]
> Or it'll help me debug my code and point out things I've missed.
I made both of these statements myself and later wondered why I had never connected them.
In the beginning, I used AI a lot to help me debug my own code, mostly through ChatGPT.
Later, I started using an AI agent that generated code, but it often didn’t work perfectly. I spent a lot of time trying to steer the AI to improve the output. Sometimes it worked, but other times it was just frustrating and felt like a waste of time.
At some point, I combined these two approaches: I cleared the context, told the AI that there was some code that wasn’t working as expected, and asked it to perform a root cause analysis, starting by trying to reproduce the issue. I was very surprised by how much better the agent became at finding and eventually fixing problems when I framed the task from this different perspective.
Now, I have commands in Claude Code for this and other due diligence tasks, and it’s been a long time since I last felt like I was wasting my time.
Have you isolated if you're properly honing in on the right breadth of context for the planned implementation?
This week I was able to tackle two long-standing bug fixes I've been noodling on and had a rough idea of what I needed to do but had competing priorities and a lack of time to sit down and really internalize the system to figure them out. I brain dumped the issue and my current thoughts and had claude formulate a plan. It solved each in less than 30 minutes of very light effort on my part. I was able to tack these onto larger work I'm doing basically seamlessly.
The other thing that I've found to be an insane benefit is filesystem-backed context switching. If your agentic workflow involves dumping your plan and progress to files in the filesystem, you can pause and restart work at any time by pointing at those files and saying "continue where you last left off". You can even take a `git diff > that-one-bug.patch` of edits made up to that point, copy that alongside the other files, and have a nice-and-neat folder of a unit of work that is ready to pick back up in the future as time permits.
I know it'll be touted as rhetoric but I have seen an order of magnitude of difference in my ability to ship things. Thankfully I don't work for a large enterprise so I don't have a multi-million line codebase to contend with or anything like that. I also, thankfully, ship projects using languages and libs that are very well represented in LLM corpuses, like TypeScript, NextJS, Postgres, though I have also found a lot of success in less popular things like Neo4j's Cypher.
I also have been massively enabled to do lots more 'ops' stuff. Being a pretty average full-stack eng means I have no experience of running sys/ops monitoring systems but LLMs only recently helped me with a bunch of docker-routing issues I was having, teaching me about Traefik, which I'd never heard of before.
Side-point: I have felt so grateful to these LLMs for freeing up a bunch of my brain space, enabling me to think more laterally and not relying so much on my working memory, severely limited now due to historic brain injury. Often people forget how massively enabling these tools can be for disabled people.
We need more empirical evidence. And historically we’re really bad at running such studies and they’re usually incredibly expensive. And the people with the money aren’t interested in engineering. They generally have other motives for allowing FUD and hype about productivity to spread.
Personally I don’t see these tools going much further than where they are now. They choke on anything that isn’t a greenfield project and consistently produce unwanted results. I don’t know what magic incantations and combinations of agents people have got set up but if that’s what they call “engineering,” these days I’m not sure that word has any meaning anymore.
Maybe these tools will get there one day but don’t go holding your breath.
That was true 8 months ago. It's not true today, because of the one-two punch of modern longer-context "reasoning" models (Claude 4+, GPT-5+) and terminal-based coding agents (Claude Code, Codex CLI).
Setting those loose an an existing large project is a very different experience from previous LLM tools.
I've watched Claude Code use grep to find potential candidates for a change I want to make, then read the related code, follow back the chain of function calls, track down the relevant tests, make a quick detour to fetch the source code of a dependency directly from GitHub (by guessing the URL to the raw file) in order to confirm a detail, make the change, test the change with an ad-hoc "python -c ..." script, add a new automated test, run the tests and declare victory.
That's a different class entirely from what GPT-4o was able to do.
I was decomissioning some code and I made the mistake of asking for an "exhaustive" analysis of the areas I needed to remove. Sonnet 4.5 took 30 minutes looking around and compiling a detailed report on exactly what needed to be removed from this very very brownfield project and after I reviewed the report, it one shot the decommisioning of the code (in this case I was using CLaude in the Cursor tooling at work). It was overkill, but impressive how well it mapped all the ramifications in the code base by greping around.
I have a disclosures section on my blog here: https://simonwillison.net/about/#disclosures
here is an example of mostly automated work. It's a small feature but it was done perfectly
The skill ceiling is high, it turns out. It's just deceptive, because it's so easy to get going. Ultra accessible foot gun, lots of work to point it in the right direction reliably and repeatedly. Significant benefits of you manage though.
I've gotten more relaxed about it now though. People will either get it or they don't.
No, I just put in less effort to arrive at the same point and do no more.
Doc comments for example. Today I had it generate doc comments for a class I wrote. I had to go and fix every single one of them because it did some dumb shit, but it out all the scaffolding in place and got the basics there so it was a lot quicker.
I also used it to generate json schemas from Python a couple of Python classes the other day. Highly structured inputs, highly structured output, so there wasn't much for it to fuck up. Took care of the annoying busy work I didn't want to do (not that schemas are busy work, but this particular case was).
Still haven't seen a use case that justifies the massive cost, or all the blatant theft and copy right infringement, or the damage to the environment...
It's like having a faster car with a bigger engine. Big deal. I want a faster car with a smaller engine. My ideal is to actually go home and stop working at the end of the day.
I also don't want to use it for my day job because I'm afraid my brain will atrophy. You don't really need to think when something is already done for you. I don't want to become someone who can only join together LLM output. I don't feel like I'll miss out on anything by not jumping on now, but I do feel like I'll lose something.
However, it's not the "coding" part of "vibe coding" that is the (semantic) problem, it's the "vibe" part.
Replacing "coding" with engineering, but keeping the "vibe" part, still conveys the meaning of not really knowing what you do, but now in an even larger surface area.
It's a lot more boring, but I'm fine with "AI-assisted software engineering" as the other-end-of-the-spectrum name opposite of "vibe coding".
Maybe something like "intent-driven development" or "AI-paired software development" might fit better?
The more accurate phrase would be "tool" or "machine-assisted development".
We don't need new terminology to describe something that has always existed because the tools we use have (arguably) improved.
"Vibe coding" is also a misnomer. People who engage in that activity are not "coding". They're using a tool to generate software until they're satisfied with the result. That has also been possible with low-code, no-code, website builders, and other tools. The more accurate term for that would be "machine-driven development".
AI-assisted coding is an absolute game changer, for me, but I would think for everyone who can wield it well. I feel I can direct a (sort-of) small army of junior coders, architects, qa- and req engineers, etc., way different than with e.g. CASE-tooling. It requires creativity, and lots of knowledge & experience in the whole software-development life-cycle to get it well, and it can adapt to any flow you like. That's really new and unique, and way more flexible than the rigid CASE-tooling that was already out there.
"Tool" seems to broad: "What are you doing?" "I'm tooling." :)
Am not highly opinionated on if "code" should be in there. The AI tools do generate code, where some low-code platforms and the likes might not. I guess "development" works as well, although "vibe developing" doesn't really have the same ring to it :)
I do like to be able to differentiate between people generating code but don't care what happens under the hood, and software-professionals that incorporate AI assisted development into their daily work on any level.
Also reviewing LLM generated code is way more mentally draining and that’s with just one agent. I can’t imagine trying to review code from multiple agents working in parallel.
Finally I’ve shipped a bunch of features to prod with Claude writing all the code. Productivity definitely went up, but my understanding of the codebase dropped fast. Three months in I’m actually struggling to write good prompts or give accurate estimates because my grasp of the code hasn’t really grown (code reviews only help so much). Curious if anyone else has run into the same thing?
I've used GPT to rapidly get up to speed with just about every aspect of circuit design, CAD, CNC... the list is long. Coding is often involved in most of these domains now, but if everything is assumed to be code-first, it leaves people who are doing different work with a constrained and apparently shrinking adjective namespace.
I'm now imagining me dying as a vibe-engineered truck has a steering/brake failure and crashes into me sending flying through the vibe-engineered papier-mâché bridge guardrails, and feeling sweet sweet release as I plummet to my doom.
Look, if you enjoy calculating a table of dozens of resistor value combinations for a feedback network that prefers reels you have on your PnP, you keep knocking yourself out.
In the example I cited, verifying a ratio isn't the hard part. It's running the dozens of permutations (smart) or hundreds of permutations (naive) that an LLM can do in 90 seconds that saves me hours of boring work. It's actually so repetitive that I'm likely to have made the same kind of mistakes you're alluding to.
As always, I end with encouragement: if you want to do everything the long and hard way, I'm not here to change your mind. You will have to stop being upset that others are moving much faster than you, though. It's a choice.
You have not met my cow orkers...
Don't get me wrong, I started vibe coding after reading Karpathy's post. I got the memo - don't review every line of code, don't stop it when it stumbles, let it recover on its own, trust the process.
But after some experience I realised I need to Constrain the model, it is like a karting track, those tires put around it keep the carts inside and safe. It's our job to set the constraints. So maybe it's "constrained agent work" not "vibe coding".
I go as far as saying the constraint harness around the agent is the new code, we can remove the code and regen it from the constraint harness and docs. What matters now is to build tests and constraints around AI work.
Anything with the word “vibe” in it sounds silly and unserious imho. What’s wrong with something neutral and descriptive like “LLM-assisted programming”? Not catchy enough?
Vibe engineering would mean giving access to Google Computer Use API to your engineering design software and letting it design the industrial component you're working on, without even looking at what it does and just telling it to "fix it" if it seems bump into problems.
> I’ve tried in the past to get terms like AI-assisted programming to stick, with approximately zero success. May as well try rubbing some vibes on it and see what happens.
Especially a term which comes across as so demeaning and devaluing to engineers (like me and yourself!)
I absolutely do not want my non-engineer friends and colleagues think I am "vibe engineering", it sounds trivial and dumbs down the discipline.
I personally believe being an engineer of some kind requires work, learning, patience, discipline, and we should be proud of being engineers. There's no way in hell I would go around and saying I'm a "vibe engineer" now. It would be like going around and saying I'm a vibe architect! Who would want to live in a skyscraper designed by a "vibe architect" ??
“casing”?
Getting good results out of that team is hard, because the bottleneck is how quickly you can review their workflow and point them in new directions.
Understanding techniques like TDD, CI, linting, specification writing, research spikes etc turns out to be key to unlocking that potential. That's why experienced software engineers have such a big advantage, if they choose to use it.
First, and most important, I have actually started a number of projects that have only lived in my head historically. Instead of getting weighed down in “ugh I don’t want to write a PDF parser to ingest that data” or whatever, my attitude has become “well, why not see if an AI assistant can do this?” Getting that sort of initial momentum for a project is huge.
Secondly, AI assistants have helped me stretch outside of my comfort zone. I don’t know SwiftUI, but it’s easy enough to ask an AI assistant to put things together and see what happens.
Both these cases refer almost necessarily to domains I’m not an expert in. And I think that’s a bigger factor in side projects than in day jobs, since in your day job, it’s more expected that you are working in an area of expertise.
Perhaps an exception is when your day job is at a startup, where everyone ends up getting stretched into domains they aren’t experts in.
Anyways, my story is, of course, just another anecdote. But I do think the step function of “would never have started without AI assistance” is a really important part of the equation.
1. Learning curve: Just like any skill there is a learning curve on how to get high quality output from an LLM.
2. The change in capabilities since recent papers were authored. I started intensively using the agentic coding tools in May. I had dabbled with them before that, but the Claude 3.7 release really changed the value proposition. Since May with the various Claude 4, 4.1 and 4.5 (and GPT-5) the utility of the agentic tools has exploded. You basically have to discard any utility measure before that inflection point, it just isn't super informative.
Likewise real world architects have the skills to design a building but do not care or know how to build it, relying on engineers for that.
I think it’s important to distinguish because we’re increasingly seeing a trend towards final product over production, meaning these “vibe” people want the tool in the end and consider the steps in between to be just busywork and AI can do for them.
That’s closer to product design than to engineering. If I can imagine Monalisa and write that thought to paper, communicating that thought and getting a painter to paint it for me does not make me Da Vinci.
Da Vinci himself likely had dozens of nameless assistants laboring in his studio on new experiments with light and color, new chemistry, etc. Da Vinci was Da Vinci because of his vision and genius, not because of his dexterity with his hands.
Clearly I’m not in marketing.
Regardless, I’m delighted that this has gotten people to ‘independently discover’ software engineering best practices on their own.
> The AI adds a ton. It really is like having a whole team of extra coders available, all of which can type faster than you.
Funny thing is, the least time consuming aspect of making programs is encoding solutions in source form. For example, a reasonable typist can produce thousands of text lines per workday if they know what must be typed (such as transcribing documents).
What takes longest when producing programmatic solutions is understanding what must be typed in the first place. After that, the rest is just an exercise in typing and picking good file/type/variable names.
I am deep in this and one important way in which managing agents is different than managing people is that micro-managing can be your friend. With human engineering colleagues, you need to allow for a healthy degree of “that’s now how I would have written the code, but it’s a reasonable way to write it.” But if my agent writes the test file in the exact same way I do, I can both review and maintain the code more easily.
I have a bunch of short markdown doc files in which I give very specific instructions for how I like code organized: much stricter than I would ever do for a colleague. I’ll tell the agent, “now add tests to this model and follow @unit_tests.md” This file specifies exactly how I like tests named, what order I like them written in the file, etc. I have docs for: models.md, controllers.md, concerns.md, and fixtures.md.
Being good at vibe coding is just being good at coding, the best practices still apply. I don't feel we need another term for it. It'll just be how almost everyone writes code in the future. Just like using an IDE.
> The developer does not review or edit the code, but solely uses tools and execution results to evaluate it and asks the LLM for improvements. Unlike traditional AI-assisted coding or pair programming, the human developer avoids examination of the code, accepts AI-suggested completions without human review, and focuses more on iterative experimentation than code correctness or structure.
How does "vibe coding" embody "best practices" as the industry generally defines the latter term?
As I understand the phrase "vibe coding", it implies focusing solely on LLM prompt formulation and not the specifics of the generated source.
> It'll just be how almost everyone writes code in the future. Just like using an IDE.
The flaw with this analogy is that a qualified developer does not require an IDE in order to be able to do their job.
Having someone cook my dinner's ingredients is just (me) being a good cook ...
If this sort of term is adopted we are in 'literally not being literally' territory.
Fortunately, I imagine that engineering outside software already has people vibing physical infrastructure solutions and call that vibe engineering so I don't think it will stick.
I don't have a great alternative unfortunately. I've used "agentic coding" in the past as a means of differentiating from vibe coding, but I don't think that's necessarily clear enough either.
That said, maybe with defining this new approach it's just going to take time for any term to become familiar enough to be self-explanatory.
Funny, because when I saw that phrase I immediately thought, "They're using chatbots to design bridges now?"
To me `vibe engineering` sounds like shoddily AI-designed suspension bridges. But then maybe I'm just an old fart programmer who thinks 'software engineering' is a term made up by programmers wanting to be as cool as bridge designers...
Here are some of my own high-level experiences / thoughts:
- Perhaps contrary to popular belief I think vibe coding will bring the best software / system architects. This is due to massively shortened feedback loop between architectural idea and seeing it in action, easiness with which it can be changed, and the ability to discuss it at any moment.
- We’re not really coding anymore. This is a new role, not a role of a senior dev reviewing PRs of junior devs. Devs are just best suited (currently) to take on this new role. I came to realization that if you’re reviewing all generated code in detail you’re doing it wrong. You just shifted bottleneck by one step. You’re still coding. You should skim if the code is in line with your high-level expectation and then make LLM maintain an architecture doc and other docs that describe what and how you’re building (this is the info you should know in detail). You can do audits with another LLM whether the implementation is 100% reflecting the docs, you can chat with LLM about implementation at any moment if you ever need. But you should not know the implementation the way you know it today. The implementation became the implementation detail. The whole challenge is to let go of the old and embrace and search for efficiency in the new setup.
- Connected to the above: reading through LLM outputs is a massive fatigue. You are exhausted after the day, because you read hundreds of pages. This is a challenge to fight. You cannot unlock full potential here if you aim at reading and reviewing everything.
- Vibe coding makes you work on the problem level much more. I never liked the phrase “ideas are cheap”. And now finally I think the tides will turn, ideas are and will be king.
- Devil is in the detail, 100%. People with ability to see connections, distill key insights, communicate and articulate clearly, think clearly, are the ones to benefit.
Hope this is helpful for others.
Looking at vibe coding: it suggests you're coding but you only vaguely know what's going on, so the work is the same (coding) but the outcome may or may not be what you want
Why dont we flip it around? We want a term that suggests that a fixed amount of work (coding) to be more efficient/leveraged.
So why dont we call it something like hypercoding, dense engineering, autocode, ...
Prompt Engineering / Directed Prompting
Architectural Steering
Spec-Driven Development
Intent-Based Coding
Critical Synthesis
Iterative Refinement
Guided Iteration
Hypothesis-Driven Coding
AI-Assisted Engineering
Cognitive Pair Programming
Dialogic Programming
Structured Prompting for Code (SPC)
Code Shepherding
AI-Assisted Engineering is probably the most descriptive. My favourite are Critical Synthesis and Code Shepherding. Both abbreviate to CS
The problem with every single tool in the category that I've come across (e.g. Conductor, Sculptor) is that they assume a single repository. Very rarely in my career working on enterprise software have I been in a situation where all my work was constrained to a single repo. Usually a story or feature spans several repos (whether split between frontend/backend, or a repo-per-service). As an engineer in these situations I never work in isolation considering only one repo -- I work across all of them at once and then prep all the PRs together. I'm not saying this multi-repo approach is good, just that it is the state of the world right now in many cases.
So imo tools like this need to work at a level above a single repo. The agent needs to start by cloning all repos needed for the task, and go from there.
Then your primary project (from Sculptor's perspective) is simply whatever contains the devcontainer / dockerfile that you want it to use (to pull in all of those repos)
It's still a little awkward though -- if you do this, be sure to set a custom system prompt explaining this setup to the underlying coding agent!
(I'm a founder of Imbue, the company behind Sculptor: https://imbue.com/sculptor/ )
You can also use GitHub submodules to implement a pattern like this, but I don't really trust them for some reason.
Thanks for your feedback! I faced this in the past. As you mentioned, monorepos are more common these days, but multi-repo is an established approach in many teams. The way I "solved" this situation was to move all the related projects into a single folder with a parent AGENTS.md file (CLAUDE.md, etc.). Then, I run Rover / Claude / Gemini on this folder.
However, this is not ideal. Due to the amount of code, it usually misses many things to do. We are currently exploring specific workflows for these use cases, trying to help agents to prepare a complete plan.
Another similar case we are working on is to support spawning the same task across different repositories. This would help teams to apply refactor or changes in different projects at the same time.
”Vibe coding” is the better term and actually makes sense for what it describes.
Leave ”engineering” in terms of taking responsibility for what you ”engineer” strictly to human professionals. That’s what people pay for and that is what makes it valuable.
It's utterly ridiculous. It feels like the PMs and higher-ups have no idea how much tech debt we're creating right now. For the past few weeks, going to work has felt like going back to school, everyone's showing off their "homework", and whoever has the coolest vibecoded instructions.md or pipeline gets promoted.
I'm slowly burning out, and I was one of the people who actually liked the technology behind all this.
Meanwhile, I have a client project where my counterpart is definitely senior to me and excitedly shares how AI is helping them solve novel problems each week!
I am currently kind of an anti-AI black sheep in engineering department because I refuse to fully embrace the exponentials and give in to the vibes.
I avoid burnout by simply switching off my brain from all this noise about vibe coding - i have thought hard and long, i know the way this is being implemented is wrong, i know they will create problems for themselves down the road (they already have, the signs are already there), i will be here to dig them out when the time comes.
So far I don't see anyone shipping faster or better with AI than I can manually, so I'm good.
Feel like there's more to this point (plus the documentation). By now we've all see the inverted-U-shaped performance curve when coding with LLMs within a single thread / context: the model starts strong, plateaus as the context fills up, and eventually declines as semantic drift sets in.
Sometimes in this situation, it works better to throw away the current code and re-implement, using what's been learnt so far, than continue working with the current context of having new contexts try to salvage the current state of some code.
Documentation is great as a reflect of the current state of some code but I've had good experiences "re-constructing" the steps taken to arrive at a current implementation by writing commit messages that are intended for an LLM, extract that later, have an LLM use it as a basis for writing a fresh "spec" for some code, that yet another LLM uses to actually write that code.
Git history is a natural place to store that...
Now everyone has examples of code they’ve “written”, but nobody can explain what it does. Unless of course, their readme.md was also completely generated.
I agree with some of the people here that vibe engineering has completely deflated my long successful career as a SWE, and it’s pushed me mentally into non-tech roles to feel motivated.
If compiles take ten minutes then sure, you’re going to want to do something else while waiting. If they take two seconds, you stop caring about keeping the machine busy because it’s impossible. It’s perfectly normal for computers to be idle waiting on you most of the time.
Great post. I've thought about what I want for better Vibe Engineering:
Each agent needs to deliver a fully working App/URL/Build so the functionality can be tested & verified.
Today AI IDEs deliver the diff + an explanation. Excellent. But give me the full build, a build I can share. A build that represents what it would be like shipped. When it comes to user facing functionality, a real build is how product owners verify a feature is complete.
Learn from Vercel -
A key part of Vercel’s magic is the automatic deployments for each branch. When working on a project with per branch vercel deployments - a team gets the immediate value of:
Shareable work - now others can see/test/give feedback on the great new feature you’ve developed - and their only work is to click a link (not git pull a branch and attempt to run locally)
No more “it works on my machine”. It either works or it doesn’t.
Confidence that if released, you know exactly what the user will experience. Give me automatic deployments for each agent, for each PR. And keep them available to spin up / re-use later.
I want to be able to re-launch it 3 months later and it just works. The reason we don’t do this today is the cost of the engineering - but with docker et al + AI agents, the cost of the eng work drops 99%
Deliver the deployment in such a way that immediate feedback to the AI could be given. This way minor tweaks can be executed immediately by the AI meaning that I can just wait for the minor tweak, review and then merge. This means the PR gets shipped NOW.
Once tooling becomes reliable and predictable enough, and the quality of the output consistent enough, using it is not a leap. Early compilers had skeptics, and GCC still has some bugs [1]
1. https://bugs.launchpad.net/ubuntu/+source/gcc-8/+bug/2101084
> Without tests? Your intern might claim something works without having actually tested it at all, plus any new change could break an unrelated feature without you realizing it. Test-first development is particularly effective with interns that can iterate in a loop.
Vibe engineer? No, try technical manager.
I know this paragraph is supposed to be encouraging, but it makes me wonder again what the actual goal of this entire AI enterprise is supposed to be.
"Less work" or "easier work" would make superficial sense, but in a society where people are in constant competition and derive both their self worth and their basis for living from work, both are effectively anti-goals. And so we get articles like this trying to offer comfort by saying that work will still be draining and challenging in the future.
So if not less work, then "more productivity", i.e. we can produce more software in a shorter amount of time (but with the same mental load). But as others have said, this was never the bottleneck.
In Claude Code for example I define a research sub-agent and let it do the majority of "research" type tasks. Especially when the research is tangential to what ever my objective is. Even if it is critical, I'll usually ask to have it do a first pass.
Script Kiddie != Hacker
Wrench Jockey != Mechanic
Vibe Coder != Engineer
It's intellectual slop. It will get the job done (atleast for a while) but without the actual growth that comes along with it. When I use an AI to one-shot a "small one-off script" I don't learn anything from it (when as a relatively new developer I SHOULD be learning something from it) And this is unlike stack overflow or googling becuase you can turn off your mind, just become one of those drones from Wall-E.
I make a point to avoid using any AI for coding (even for looking things up) when working on personal projects, at the cost of "productivty" and "efficiency" , but I get to retain my humanity and soul in programming.
Sorry if this sounds cheesy, it's just I care deeply about code craftsmanship from my end, to see that skill be diminished to an random number generator? Yeah No.
Knowing when the agent has failed and it’s time to roll back. After four or five turns of Claude confidently telling you the feature is done, but things are drifting further off course, it’s time to reset and try again.
One of the most underrated skills in effectively using gen-AI for coding is knowing ahead of time whether it will take longer to carefully review the code it produces, versus writing it from scratch yourself.
[0] https://www.joelonsoftware.com/2000/04/06/things-you-should-...
[1] https://mattrickard.com/its-hard-to-read-code-than-write-it
[2] https://trishagee.com/presentations/reading_code/
[3] https://idiallo.com/blog/writing-code-is-easy-reading-is-har...
[...] https://www.google.com/search?q=it%27s+harder+to+read+code+t...
Agentic coding is just doing for development what cloud computing did for systems administration. Sure, I could spend all day building and configuring Linux boxes to deploy backend infrastructure on if the time and budget existed for me to do that, and I'd have fun doing it, but what's more fun for me is actually launching a product.
For instance, I recently had to replace a hard-coded parameter with something specifiable on the command line, in an unfamiliar behemoth of a Java project. The hard-coded value was literally 20 function calls deep in a heavily dependency-injected stack, and the argument parser was of course bespoke.
Claude Code oneshotted this in about 30 seconds. It took me all of 5 minutes to read through its implementation and verify that it correctly called the custom argument parser and percolated its value down all 20 layers of the stack. The hour of my time I got back from having to parse through all those layers myself was spent on the sort of programming I love, the kind that LLMs are bad at: things like novel algorithm development, low-level optimizations, designing elegant and maintainable code architecture, etc.
For more complex modifications, I would have taken the time to better internalize the code architecture myself. But for a no-brainer case like this, an LLM oneshot is perfect.
It's not so trivial to verify that the change doesn't cause problems elsewhere, where it also should have been propagated.
You raise a good point: an important skill in effectively using LLMs for coding is both being able to recognize ahead of time that cases like this are indeed simple, but also recognizing after the fact that the code is more complex than you initially realized and you can't easily internalize the (side) effects of what the LLM wrote, warranting a closer look.
It is very easy to notice at work who actually likes building software and wants to make the best product and who is there for the money, wants to move on, hard code something and get away with the minimal amount of work, usually because they don't care much. That kind of people love vibe coding.
Today: 'code is cheap, show me the talk'
Not much to base that on other than vibes, though :)
Instead of just vibing something out, pushing it to prod and seeing the problems. Or not even checking...
I can see a plausible future where if we go down this route, what I call coding right now will feel the same.
For example if someone creates a Pull Request "#2122 feat: User password change"
Automation would deploy it to pr_2122_feat_user_password_change.mycompany.com
Heroku: https://devcenter.heroku.com/articles/github-integration-rev...
Vercel: https://vercel.com/docs/deployments/environments#preview-env...
Don’t mind me, I’m just vibing.
Agentic coding, inner loop, the stuff an engineer does on his laptop. Agentic engineering, outer loop, sdlc across the organization
Not perfect, but good enough.
It’s also why these tools feel great in green field projects, since these ones don’t typically have any code review in place (i.e. one dev going brrr)
I was a solo dev going brrrr with no code review, using Cursor to make pretty sweeping changes and implement full features. I think because it was a fork I was able to go really fast because I could point the coding agent at existing patterns and say, "do this but different".
Occasionally I would test changes in staging but mostly would just promote them straight to production. I went on like that for ~3 months and the product ended up getting ~75k (daily!) users in that time with *no* outages or downtime (!!!)
It worked really well! Until I hit my scaling limit and we added more engineers to the team. A new hire on their second week on the job shipped a change that was also written with AI and caused a 7 hour outage before we realized something was wrong. I even *thoroughly reviewed the code* because I knew it was generated and I made a comment questioning the lines that caused the outage, but after we talked it over we decided it was fine and was only one tiny bit of a much larger change.
So I guess, AI will ship bugs but so will peer-reviewed code.
Yes, PRs start piling up because devs can vibe code them faster than they can be competently reviewed. This is a problem with the vibe code process, not the code review process.
I want to reach out since I want to express my disappointment regarding this thing you write here:
"I propose we call this vibe engineering, with my tongue only partially in my cheek."
Back in march 2025, you and I had a brief convo on X about this very term that came to my mind "Vibe Engineering":
https://x.com/pierre_vannier/status/1904933441042317821
I also mentioned it in our podcast in April 2025: https://x.com/Flint_company/status/1909946181897044195
Yesterday, I pinged you on X about this "term's paternity" in your thread without any response: https://x.com/pierre_vannier/status/1975579806495293910
You and I met in AI Engineer in SF last June and we discussed for 1 hour (not about this specifically but about AI and all).
At the very least, it could be cool to mention it in your post or newsletter since, I do think this little convo on X + our talk in SF might have just planted the seed in your head for this "invention" of yours...
I don't want to stay in history books but I think it's cool to also credit people when you iterate on their idea instead of just make the thing entirely "yours"...
Best Pierre
If you're fine with an AI designed bridge, where the builder tweaked the prompt to get the LLM to suggest cutting out all safety features, then okay, go with vibe engineering.
this sadly seems to sum up most of this new wave of work. hopefully we can find a better workflow that still feels as good as coding used to
* There are actual productivity gains to Agentic Coding (my tool of preference is Claude Code)
* There is a big chance of producing a lot of slop if one does not use it with care
* More than ever checking mechanisms are essential (static type checking and a good unit testing suite)
* Everything that was boring and tiresome about unit testing is now mandatory since you will be writing almost none of it.
>> Supporters see it as a bridge between humor and professionalism >> Critics hear it as the sound of their trade being rebranded into a meme.
Both sides, though, agree on one thing: coding is no longer the same craft it was a year ago.
My opinion: I believe “vibe coding” has already broadened to mean any AI-assisted coding, making “vibe engineering” redundant.
[1] https://simonwillison.net/2025/Jun/27/context-engineering/
That's more than $5,000 a month assuming 40 hours a week.
The end result didn't compile by the way.
To add as a data point: I've met founders that are business founders and are vibe coding entire businesses with React, Supabase and Netlify.
LLMs allow for amazingly high fidelity interaction designs to receive funding and prove out certain PoCs. Whether his stuff will stay up in production remains to be seen.
The part about past management experience being a key skill surprised me but now it makes sense.
I really do usually have 3 different projects in flight for at least 6 hours a day. I'd write a blog post but I keep expecting someone else will write the essential same post tomorrow. :)
p.s. The first 2 months was exhausting but now it's slightly less exhausting. Make no mistake, it is an extreme transition to make.
For the experienced lot of us, I've heard many call it "hyper engineering"
The AI agent in Cursor with Gemini (I'm semi-new to all of this) is legit.
I can try things out and see for myself and get new ideas for things. Mostly I just ask it to do things, it does it; for specific things I just highlight it in the editor and say "Do it this way instead" or "for every entry in the loop, add to variable global_var only if the string matches ./cfg/strings.json" I _KNOW_ I can code that.
But I like my little clippy.
And I think that is actually the most natural meaning for "vibe engineering", actually: Parallel to "vibe coding" where you serially prompt the AI to write the code for you, "vibe engineering" should be serially prompting the AI to do the entire engineering process for you.
I also predict that a precisely defined term for what Simon is describing will inevitably end up being primarily used by people who are actually doing "vibe engineering". Being disciplined and careful is hard.
People and organizations love to claim/pretend they're doing expensive, mostly invisible work like "building secure software". Given nearly every organization claims they use security best practices no matter what their actual practices, I imagine it will be that way with "actually reading and verifying what the LLM generated for me".
Certainly I've been disappointed how often someone presents something they've "written" in recent months that turns out, on inspection, to be AI slop that the "author" hasn't even read every sentence of carefully.
Not because I disagree with you, I don't! It's because I fear a gradual brain drain of people who actually love their craft and know how to build things properly. I fear we'll end up with worse software thats simply 'good enough', built a atop a pile of AI slop.
If it's cheaper but with acceptably worse results, I fear this is good enough for a lot of companies.
We didnt stop calling them Framers or Finish Carpenters when they got electric saws and nail guns.
Tooling does not change the job requirements.
If using LLMs makes you slower or reduces the quality of your output, your professional obligation is to notice that and change how you use them.
If you can't figure out how to have them increase both the speed and the quality of your work, you should either drop them or try and figure out why they aren't working by talking to people who are getting better results.
My largest project is a year old, it's full-stack JavaScript, and I consciously use patterns, structures, and diligently add documentations right from the beginning for the code base to be as LLM friendly as possible.
I see great results on refactoring with limited scope, scaffolding test cases (I still choose to write my own tests but LLMs can also generate very good tests if I explicitly point to existing tests of highly related code, such as some repository methods), documenting functions, etc. but I'm just not seeing the kind of quality that people claim that LLMs can do for them on complex tasks.
I want to believe that LLMs are actually capable of doing what at least a good junior engineer can do but I'm not seeing that in my own experience. Whenever we point out these issues we are encountering, we just basically get the "git gud" response with no practical details on what we can actually dp to get the results that people claim to be getting. Then people start blaming our lack of structures, patterns, problems with our prompts, the language, our stack, etc. when we complain about the "git gud" response being too vague. Nobody claiming to be seeing great results seems to want to do a comprehensive write-up or, better still, a stream of their entire workflow to teach others how to do actual, good engineering with LLMs on real-world problems either -- they all just want to give high level details and assert success.
On top of that, the fact that none of the people I know in engineering working in both large organizations and respectable startups that are pushing AI are seeing that kind of results naturally makes me even more skeptical of claims of success. What I'm often hearing from them are mediocre engineers thinking that they are being productive but actually just offloading the work to their colleagues through review, and nobody seems to be seeing tangible returns from using AI in their workflow but people in C-suites are pushing AI anyway.
If just about anything can be "your fault", how can anyone claiming that LLMs are great for real engineering without showing evidence be so confident that what they're claiming but not showing is actually the case.
I feel like every time I comment on anything related to your blog posts I probably came across as belligerent and get down voted but I really don't intend to.
The only issue with "Handled Programming" is I don't like how it fits for a name.
Vibe is much too unserious to pass my check for a way of professionally doing something and it also does not reflect the level of engagement I have with the AI and code since I'm putting together specs and otherwise deeply engaging with the model to produce the output I want.
The reality is the tools are really useful when used as tools, like a power drill vs. a screw driver.
Vibing implies backseat driving which isn't what using the tools proficiently is like. The better term would be 'assisted' or 'offloaded'.
Same thing with the term 'engineering'. That's a fairly new term that implies being engineers which we are not. We haven't studied to be engineers, nor have real engineering degrees. We've called ourselves that because we were doing much more than the original job of programmer and felt like we deserved a raise.
'LLM extended programming' is not as catchy but more relevant to what I observe people doing. It's valuable, it saves time and allows us to learn quicker, very powerful if used properly. Calling it 'vibe engineering' is a risky proposition as it may just make people's eyes roll and restrict us to a lesser understanding.
Uh, speak for yourself. There are countries where being a software engineer does indeed imply that you studied engineering and hold a "real" engineering degree.
Also, Hillel Wayne's "Are We Really Engineers" is worth reading:
As "coders" or "programmers", some of us should answer the question "are you an engineer?" with a proud "of course not!" (That's me.) And some of us should answer, equally proudly, "of course I am!"
Hillel Wayne's series is great.
I have fairly decent engineering credentials, but when the task fits, I prefer to vibe code.
Just don't capitalize it in Oregon.
I mean, the AI companies have an incentive to make it easier, right? The next iteration of coding agents will be better at working with their human PHBs to divine intent and turn vague specifications into something workable - a key skill of human engineers.
Is this really any different from how Google learned to understand search queries?
Why would you add “vibe” then? Seasoned devs don’t vibe, they know what they do
This is making me like the written-by-math-grad-school-interns python mess I inhereted!
Ill stick with writing code and just use AI for snippets/stackoverflow replacement.
Vibe coding is different because it's the "dictated but not read" of coding. Yes, I was around when the LLM was writing the code, and I vaguely instructed it on what to write, but I make no assurances on the quality of the output.
I'd offer a new term: curmudgeon coding. This pre-dates LLMs and is the act of engineers endlessly clutching pearls over new technology and its branding. It's a reflexive reaction to marketing hype mixed with a conservative by default attitude. Think hating on "NoSQL". Validity of said hate aside, it's definitely "a type" of developer who habitually engages in whinging.
Seriously? The term has been around in the community for as long as "vibe coding" has been around. While both sound equally cringe, the statement leaves a bad taste.
That’s so much better than wasting time on the frustrated who just negate without a meaningful try.
I share most of the experience and learning with the author, I just still don’t know how to name the whole process.
The whole vibe thing has already brought negativity into terminology because of all negating.
Closest thing in history of software engineering practice from the past is XP (Xstreme Programmimg) with its pair programming approach. It was a precursor of anything modern agile. Invented at the end of 90s
It’s just this time, my pair programming mate is computer instead of human person, but I treat it as junior to me and review, organize, while coding, testing, documenting is delegate and we jointly do the analysis as a part of the discussion.
I can agree, it’s strongly augmented experience and weird often to newcomers to adapt, but when a critical path to succeed is discovered, it works great and results are a blast!
Seriously now, I think the whole industry suffers from too many buzzwords and whacky terminology.
*The job hasn't changed*. As mentioned, all those things from the past are still the most important thing (version control, being good at testing, knowing when to outsource, etc).
It's just coding (which is something that was never about typing characters, ever).
Think this might be a typo
Software engineer.
Not vibe engineer. Not Java engineer. Not compiler engineer.
You either understand the computer and then can apply any fucking tool or methodology to solve a problem, or you’re out.
I was hoping that "vibe engineering" was going to be designing g bridges in the same way people think they can build apps with vibe coding
That would be alarming
1. It is confusing in the sense that all other engineering disciplines have the form "<X> engineering" where X is the object being engineered. This makes it sound like we are engineering vibes, not software. 2. Software engineering was already only marginally worthy of the term "engineering". There is a strong subset of computational system building that I think actually meets the standard of engineering (rigorously understanding requirements and system bounds, probably showing that the system achieves stability within those bounds). This just makes that situation worse. The devil is in the details, and over-reliance on llms ultimately makes those details a black box to the designer. Before you claim it's the same as when a designer relies on a suite of humans—no. The designer can ask those humans for reverentially transparent proofs that certain conditions are upheld, there is a social accountability structure, etc etc all of which does not exist with LLMs even if they get good enough that we can just trust them.
If we are going to keep calling software construction engineering, can we please at least get industry wide standards around ethics before we go off vibing? There's so much that we still sorely need in order to really make this discipline rigorous and socially accountable in the first place.
It's high time software engineers do the same if they really want to be engineers and not just corporate lackeys who happen to know how to program and/or design computational systems.
I am on the stupid side, personally.