Agentic Engineering is just Vibe coding
Posted by dark_mode_everything@reddit | ExperiencedDevs | View on Reddit | 78 comments
Yet another AI post but hear me out. My team consists of about 8 very senior devs and some of them (especially one guy) very heavily uses AI. They insist that they're not vibe coding and that they review the generated code so it's "AI assisted engineering" or "agentic Engineering" or some bs like that.
But imho just reviewing (or understanding each change in isolation) isn't enough. Unless you run every code path and manually evaluate each line of your code you're pretty much vibe coding. And what these people, even as seniors, don't understand is that more code is not a good thing. AI just creates mountains of it and then everyone needs to spend more time managing it. They no longer spend time thinking of a nice and simple (or elegant) solution but the first thought is hey let's ask AI. It's infuriating.
Apologies, rant over.
ub3rh4x0rz@reddit
I feel dumber for having read this and every post like it. Write in a diary. This is not earnest, and it causes second hand embarrassment. This isn't 2023 chatgpt era dude.
dark_mode_everything@reddit (OP)
Well.....I am an old man. And I am yelling at clouds.
ub3rh4x0rz@reddit
You're reacting to fear rooted in change and slandering your peers to make yourself feel better. You should have accrued more self awareness by now.
dark_mode_everything@reddit (OP)
Based on the number of comments you made here I reckon I could say the same about you. You're probably reacting to fear rooted in your over reliance on ai and want everyone else to be like you. Or who knows? We're both making assumptions here, aren't we.
Also, about your other comment, Linus used AI for his fun side project. He said he would never use it for anything serious. Lol.
ub3rh4x0rz@reddit
No, I'm sick of hearing dates and just generally trite takes like yours. I started using claude code relatively recently A (beginning of the year like most), B, still write code by hand at least 50% of the time, and C, you're foolish if you literally think people become unable to write code after a handful of months (just goes to show your fear based position).
Re Linus, Linux allows AI assisted commits. Linus did say that, prior to Linux allowing AI assisted commits. His position has obviously evolved.
dark_mode_everything@reddit (OP)
So many mentions of the fear thing from you it feels like projection...
ub3rh4x0rz@reddit
You're the one burying your head in the sand and painting with comically broad strokes. The silent majority of SWEs who are actively learning how to use this technology responsibly are sick of your self-serving bullshit.
Salty-Wrap-1741@reddit
This.
DocDavluz@reddit
I'm starting to grasp the difference when looking at my self appropriation of the code. Speed has the cost: you got more code to grasp in the same timebox, your brain becomes the bottleneck. I'm now more in a balanced usage, sometimes not generating code with AI for the sake of my own comprehesion of what's made.
ub3rh4x0rz@reddit
This is the way.
ub3rh4x0rz@reddit
It honestly helps maintain objectivity. 20 minutes, 2 hours, whatever, is a long time to adhere to your own design, and by the time it's over, there's really strong incentive to trust your own process. Being an adversarial reviewer of your own output is important.
teerre@reddit
If you change the meaning of a name, the meaning changes. Yes. Shocking
Let's then forget about these labels and just compare not being a subject expert in some field, telling a LLM how to superficially achieve something, not reviewing the output with being a subject expert in some field, knowing precisely how to build, test and deploy something and reviewing each line produced. It seems evident that the two are not the same
thisisafullsentence@reddit
Vibe coding means you're not reading the code. Generating the code then reviewing it thoroughly is not vibe coding.
MoreRespectForQA@reddit
Ive tried doing this and I always ended up:
review -> regenerate -> review -> regenerate -> etc. often with the last step being "fuck it I'll write it myself".
vibe coders say "ah it's a skill issue you're not prompting it right". i say theyre probably either not reviewing their output as thoroughly as they think they are or are building something pretty basic.
ub3rh4x0rz@reddit
Honestly what you described usually yields better results than just writing it yourself once and your buddy commenting lgtm.
MoreRespectForQA@reddit
not to put too fine a point on it but this is something I've found that junior programmers, mediocre programmers and non programmers all agree on.
simfgames@reddit
I didn't block anyone. I just deleted my message because I remembered I usually hate discussing things on here.
I work on complex problems using an AI workflow I've put together from scratch. I've been able to improve both quality and quantity of AI output significantly. I've been at it for over a year, full time. It has consumed my work life, and I'm intimately familiar with how it all works.
So it's usually a waste of time to discuss anything on here. I press reply, and some dipshit who has used copilot for the last month (by management mandate) shows up within 5 mins to tell me I'm a moron.
I'll pass. You're right, and all this Ai stuff is nonsense.
simfgames@reddit
If you wanna talk expertise, there are a ton of very high level programmers claiming that AI speeds them up significantly.
Based on your replies though, I'm sure you know more than John Carmack.
ub3rh4x0rz@reddit
Not to put too fine a point on it, but your little snide insinuation just makes you look like an arrogant midlevel, not a high performing senior+. I see your condescension and raise you that even the Linux kernel developers see the benefits and officially allow AI assisted commits now. Virtually every senior+ engineer is using AI daily at this point, but keep telling yourself you are some sort of unicorn.
MoreRespectForQA@reddit
kernel developers havent taken a position on whether it boosts productivity they just said that you are responsible for your code as the submitter. This is sensible position if
is there a reason you chose to misrepresent their position?
some projects have taken a stricter position due to an influx of slop and banned it entirely.
nonasiandoctor@reddit
Yeah I would argue they are accepting the tradeoff of taking AI assisted submissions because it means they will have more people participating in the Linux ecosystem. Not encouraging it.
MoreRespectForQA@reddit
It's an ecosystem that runs on trust. If you burn your trust then you're locked out as a kernel dev forever which is quite the burn. This should disincentivize the drive by slop problem to an extent.
Expert-Reaction-7472@reddit
unlike you, the snowflake writing unique and complex code.
MoreRespectForQA@reddit
just tell yourself it's a skill issue and i dont know how to prompt if im making you feel insecure.
Expert-Reaction-7472@reddit
nah im not the one claiming to have some exceptional level of coding talent that means i work on really special things while everyone is is doing some variation of internet plumbing.
ub3rh4x0rz@reddit
Unique and complex code is a consequence of bad engineering 99.999% of the time.
Expert-Reaction-7472@reddit
yep. most apps might have a tiny bit of necessary complexity which if designed well is neatly contained.
Unfortunately devs love making easy stuff hard for some reason.
ub3rh4x0rz@reddit
It has never been cheaper to produce code, review it with zero emotional attachment, and strip it down to the purest most elegant solution with the least cognitive burden for understanding it and the fewest moving parts. I like slinging brackets as much as the next gray beard, but I like making decisions that result in the best systems at least as much. Crazy times we live in.
PeachScary413@reddit
The thing is... you might tell yourself that you are actually reviewing it, and you might even believe it fully.. but you won't because eventually you get used to push the LGTM+2 and merge since "you haven't found any faults so far and it's late and I can't be bothered to review 5000 lines of slop"
ub3rh4x0rz@reddit
When you resort to cherry picking the 1% shittiest most dishonest AI usage and putting it forth as a representative example, you're basically just communicating that you're wildly out of touch. Anyone shipping 5k LOC changes is a douchebag or is under duress.
Dry_Builder_1251@reddit
Ding ding
djnattyp@reddit
slop plop
Risitop@reddit
and then you woke up
thisisafullsentence@reddit
This is not my experience.
OtaK_@reddit
That's true in a vaccuum, but there's no way people are honestly telling themselves that they properly reviewed the magnitude of code that gets output my LLMs. Agentic engineering is just delusional vibe coding with a conscience.
And if someone actually properly reviews that stuff, then there's no way they're more productive than just going ahead and writing it themselves.
Leather_Secretary_13@reddit
I'll spend an evening generating a feature and, if I'm doing a lot of back and forth, then it's easier to have it fix a single line by generating the whole file than it is to fix it myself.
If I fix it myself, the ai context doesn't know about it, and if I use tokens to just tell it what I did then it likely forgets in the next iteration.
If it's a graph algorithm, sometimes it's easier to write the main structure myself, but after that it's all ai still.
I don't really use git anymore though and hate vscode. Perhaps cursor got better since I last used it but I just found it annoying after the initial new cool facgor wore off.
ub3rh4x0rz@reddit
You don't use git anymore? Ok so you became a PM basically? Cool. I'm skeptical you're an experienced dev.
Forsaken-Promise-269@reddit
Ok I don’t agree that agentic engineering is the same as Vibe Coding
It also heavily depends on the nature of the coding problem or the type of feature you are working on
Can you name specifics of a feature they have shipped with ai coding that you feel is inadequately engineered unproven or full of issues?
Had you written the code manually and if it has issues the act of writing the code forms a mental model of how it works, so issues are usually just part of the human process that get fixed in review , testing or bug fixes
I do agree that AI coding breaks that model and also creates a lot of code - which is an uneasy feeling - rather like sitting in self driving car
If so why isn’t there a process to vet the code better, test, use agents to review code and have a more vetted integration process?
ub3rh4x0rz@reddit
When usage isn't fueled by sycophantic executives, but is adopted in a bottom up fashion, there is no reason the marginal change has to be all quantity no quality. I have a lot more time to focus on design, whether it is at the system level or at the level of the code. People tend to conflate the effects of top down corporate mandate and the sorts of environments where that occurs to the technology itself. It can be a >1.0 quality multiplier when used responsibly.
WalidfromMorocco@reddit
Honestly depends on the amount of code you are dumping with these agents. If you are generating entire features using LLMs, then I don't think any amount of code review will have you truly understand the code and the decisions that the LLM had taken, unless you are really spending hours going back and forth with it, which seems to be counter productive.
notAGreatIdeaForName@reddit
Yes, getting back the code ownership feels like fighting in a war after generating code for a large feature.
HolyPommeDeTerre@reddit
Which is weird. Writing code wasn't hard, having ownership was natural. Now we review the code of "the average person on the internet" to be sure it fits our standards. This is harder, and not faster.
Where is the gain?
dark_mode_everything@reddit (OP)
The gain is that you can outsource your thinking, and generate 10k lines in a few minutes for a feature that can be implemented with 1000 lines. /s
HolyPommeDeTerre@reddit
Glad you put the /s. Some do have this take unfortunately.
ub3rh4x0rz@reddit
This is a minority of actual working senior+ devs who use AI, which is most actual working senior+ devs. The vast majority see it for what it is, and are learning how to better utilize it every day. The next most prevalent group is frozen by fear and is resistant to engage with it, still a minority though. It can offer wild upside in some cases, minimal downside in others, and at least on par in the majority, all while experimenting conservatively. That's good enough to actually invest in learning how to benefit from it, is where most of us are at now. I get that a lot of people are dealing with this in highly toxic work environments, but my sympathy runs out when the cope is to slander their peers and frankly their betters. Nobody commenting in this thread from a never AI position is better at the craft than Linus Torvalds or most other kernel devs who have found ways to use it beneficially, it's just straight up arrogance.
notAGreatIdeaForName@reddit
The gain is the "speed" (if you don't take back the ownership and have no idea how exactly things work). If you want to do it right it is a lot harder especially because you have no natural memory map of the sytem.
The selling point has been an unfair comparison from the start and I think the industry needs some time to find out to which degree this is okay.
Currently everyone and their grandma is hyped but as token costs start to become less subsidized, the enshittification starts and some services have been suddenly less stable than they used to be we will see how this plays out. I
HolyPommeDeTerre@reddit
I was going rhetorical on my question here. My bad for not making it clear enough.
But I agree with you, the "speed" sold by the hype, when reality isn't balanced yet. When you don't actually just look at the code output but the overall health of your codebase and devs knowledge, it doesn't look good. In the end, you take energy and time to get back those.
It's like we have a marathon to run and one guy sprints to the first kilometer/mile and say, see, we can do that for the distance left.
ohhi23021@reddit
hours back and forth still produces features that normally take weeks in a few days.
ub3rh4x0rz@reddit
Or months in a couple weeks...
Or the same amount of time, but higher quality / simpler.
People are projecting, reacting to fear, and generally begging the question that AI usage is incompatible with high quality engineering. Ultimately it is all rooted in fear.
WalidfromMorocco@reddit
Sure. At work we have a client who specifically request we use genAI to develop their project to cut costs. We did code reviews and written extensive prompts so the LLM does not go astray from our vision. Now, if you ask any of us how that piece of software works, most will scratch their heads. However you can ask me questions about projects I've worked on years ago and i can still answer you and tell you why certain decisions were made.
That being said, if you tell me it works perfectly for you, I won't push back
dark_mode_everything@reddit (OP)
This is exactly my point.
tieno@reddit
slop coding?
m2ilosz@reddit
Either way reviewing the code isn't writing it. I know "vibe coding" is supposed to be derogatory, and it's often used like that, but I like having different name for this completely different action of "coding without coding".
Bricktop72@reddit
Aww shit, I've been vibe coding for 30+ years. In fact ever developer I have ever worked with is a vibe coder.
demchaav@reddit
AI coding tools are great until they confidently explain the bug they introduced three prompts ago. Then you're debugging the explanation, not the code.
muntaxitome@reddit
Vibe coding is a term where we have an original definition by the person that coined the term:
If you are thinking about the code it is not vibe coding. If you review the code, it is no longer vibe-coding (even if it started out as it). As for 'everything except evaluating every line of code is vibe coding'. Hard disagree. Then adding a library to your project or copy pasting an algorithm would also be vibe-coding?
The code process does not meet your quality standards. That is fine. But vibe coding is something else.
dark_mode_everything@reddit (OP)
Okay. Then Slop coding?
muntaxitome@reddit
Don't forget to record the reaction when you call it that to their faces.
geon@reddit
Exactly. Code is a liability, not a resource.
ub3rh4x0rz@reddit
True, but also, delete your codebase and all liabilities are eliminated, yeah?
I guarantee all of OP's senior colleaugues know this well and would say "you're not really applying that with nuance".
geon@reddit
Some liabilities are necessary. That’s how you make money.
But taking on more liabilities without a good reason is a bad idea.
That’s plenty nuanced.
Dry_Builder_1251@reddit
"And what these people, even as seniors, don't understand is that more code is not a good thing"
Seniors who measure anything anytime in LOC?
Run
ub3rh4x0rz@reddit
Is it possible that OP is experiencing confirmation bias? No, let the circle jerk commence /s
Dry_Builder_1251@reddit
Ofc it is possible.
But I don't even try to guess what is going through anyone's mind in random, anonymous, text based forum.
I cannot get it right, bc I don't have like 99% of info. So I take everything at face value: I pretend it is as said, never actually believing a word.
ub3rh4x0rz@reddit
So you take another person's speculation as to what's going through their peers' minds at face value? That's just outsourcing judgement.
Dry_Builder_1251@reddit
It is exact same if I do it or if someone else does it.
I, nor anyone else except the one saying, has any kind of idea about their situation, trains of thought, world view, etc.
Analysing others based on random text and your own prenotions can't result in anything else except misunderstandings.
So why bother.
Expert-Reaction-7472@reddit
it's so boring hearing this shit.
AI is here. People are going to use it.
Find ways to have productive conversations about that instead of complaining about how it isn't how it used to be.
ub3rh4x0rz@reddit
Even the complaining is ok from time to time if it isn't laced with self righteous and delusional "every one of my peers has lost their mind, I am the last defender of thr crsft" narrative.
Expert-Reaction-7472@reddit
...the last few NAND to tetris heros will save the world with their low level coding skills when we have to reboot society due to a series of incredibly unlikely events happening.
SansSariph@reddit
I'd encourage you to pair with the "one guy" and ask him to show you his workflow.
Maybe he's actually vibing and hitting "accept". Maybe he spends an hour workshopping acceptance criteria and a test plan and has a novel adversarial review setup that enables him to build confidence in his output.
What sort of bugs are you finding in his output?
ub3rh4x0rz@reddit
While I agree in spirit, having tried the whole "up-front robust plan and large change set", it doesn't work well that way. Plan mode yes, large change sets without review, no. Targeted commits and review before moving to the next is the only way.
SansSariph@reddit
Agreed! I'm mostly just trying to encourage some curiosity, OP might learn something. Or they're right and the shop is a slop factory, that's totally possible too.
dark_mode_everything@reddit (OP)
Not so much bugs, but unnecessarily verbose or bloated code. Code that somehow does the thing but doesn't do it in the simplest way.
ub3rh4x0rz@reddit
Is he writing sensitive business logic or centering a div? Does it matter or no?
CodelinesNL@reddit
Strong opinions from someone with very little experience.
AI is a tool, a very powerful one. For the tasks I have it does not create more code than I would by hand.
People should be required to actually post their tools and workflows when making these kinds of claims. No one cares about your shitty Copilot experience, everyone (even Microsoft) knows it's shit.
superpitu@reddit
Not quite. I use Claude heavily, but I won’t push the PR until I am happy to call that code my own. No design compromises, no shortcuts, no hacks, the code needs to be solid. There are points where Clause just doesn’t get it and ai roll my sleeves and write that bit myself. I don’t care who wrote it, but I do care that whatever I push is not slop.
notAGreatIdeaForName@reddit
> Unless you run every code path and manually evaluate each line of your code you're pretty much vibe coding.
If they don't run it and give it to review and it doesn't work at all because they didn't bother to do the slighest QA then I would be pretty pissed.
__bee_07@reddit
This is a wrong take on agentic engineering .. you are mixing things