I recently used a coding agent on a project where I was using an unfamiliar language, framework, API, and protocol. It was a non-trivial project, and I had to be paying attention to what the agent was doing because it definitely would go off into the weeds fairly often. But not having to spend hours here and there getting up to speed on some mundane but unfamiliar aspect of the implementation really made everything about the experience better.
I even explored some aspects of LLM performance: I could tell that new and fast changing APIs easily flummox a coding agent, confirming the strong relationship of up-to-date and accurate training material to LLM performance. I've also seen this aspect of agent assisted coding improve and vary across AIs.
I'm not a professional SWE, I just know enough to understand what the right processes look like, and vibe coding is awesome but chaotic and messy.
If frameworks don't make sense for what you're doing though and you're now relying on your LLM to write the core design of your codebase... it will fall apart long before you reach "its basically working".
The more nuanced interactions in your code the worse it'll do.
There is a big difference between vibe coding and llm assisted coding and the poster above seems to be aware of it.
It was already obvious from your first paragraph - in that context even the sentence "everything works like I think it should" makes absolute sense, because it fits perfectly to limited understanding of a non-engineer - from your POV, it indeed all works perfectly, API secrets in the frontend and 5 levels of JSON transformation on the backend side be damned, right ;) Yay, vibe-coding for everyone - even if it takes longer than the programming the conventional way, who cares, right?
The exuberance of rapid early-stage development is mirrored by the despair of late-stage realizations that you've painted yourself into a corner, you don't understand enough about the code or the problem domain to move forward at all, and your AI coding assistant can't help either because the program is too large for it to reason about fully.
AI lets you make all the classic engineering project mistakes faster.
It depends. No one is running their brain at full-throttle for more than a few hours on end.
If your "niggling" defects is mostly changes that don't require deep thought (refactor this variable name, function parameters/return type changes, classes, filename changes, etc), then I can see how it is energising - you're getting repeated dopamine hits for very little effort.
If, OTOH, you are doing deep review of the patterns and structures the LLM is producing, you aren't going to be doing that for more than a few hours without getting exhausted.
I find, myself, that repeatedly correcting stuff makes me tired faster than simply "LGTM, lets yolo it!" on a filename change, or class refactor, etc.
When the code I get is not what I wanted even though it passes the tests, it's more mental energy to correct the LLM than if I had simply done it myself from the first.
A good example of the exhausting tasks from today - my input has preprocessing directives embedded in it; there's only three now (new project), so the code generated by Claude did a number of `if-then-else-if` statements to process this input.
My expectation was that it would use a jump table of some type (possibly a dictionary holding function pointers, or a match/switch/case statement).
I think a good analogy is self-driving cars: if the SDC requires no human intervention, then sure it's safe. If the SDC requires the human to keep their hand on the wheel at all time because it might disengage with sub-second warnings, then I'm going to be more tired after a long drive than if I simply turned it off.
It's often that just getting started at all on a task is the hardest part. That's why writers often produce a "vomit draft" (https://thewritepractice.com/vomit-first-draft/) just to get into the right frame of mind to do real writing.
Using a coding agent to fix something trivial serves the same purpose.
At some point you realize if you want people to trust you you have to do this. Otherwise you’re just gambling, which isn’t very trustworthy.
It’s also got the cumulative effect of making you a good developer if done consistently over the course of your career. But yes, it’s annoying and slow in the short term.
You, too, can be awarded the Order of Labor Glory, Third Class.[1]
Is that contrarian though? Seems like pretty normal corparate setting bragging to me. (Note: I am not accusing you of it since your boss or collegues does not read this).
On the variant of "I am bad at not working too hard".
You didn’t find that to be a little too much unfamiliarity? With the couple of projects that I’ve worked on that were developed using an “agent first” approach I found that if I added too many new things at once it would put me in a difficult space where I didn’t feel confident enough to evaluate what the agent was doing, and when it seemed to go off the rails I would have to do a bunch of research to figure out how to steer it.
Now, none of that was bad, because I learned a lot, and I think it is a great way to familiarize oneself with a new stack, but if I want to move really fast, I still pick mostly familiar stuff.
I was ready to find that it was a bit much. The conjunction of ATProto and Dart was almost too much for the coding agent to handle and stay useful. But in the end it was OK.
I went from "wow that flutter code looks weird" to enjoying it pretty quickly.
So don't. It is vibe coding, not math class. As long as it looks like it works then all good.
AI removes boredome AND removes the natural pauses where understanding used to form..
energy goes up, but so does the kind of "compression" of cognitive things.
I think its less a quesiton of "faster" or "slower" but rather who controls the tempo
compression is exactly what is missing for me when using agents, reading their approach doesn't let me compress the model in my head to evaluate it, and that was why i did programming in the first place.
But I 100% agree. It's liberating to focus on the design of my project, and my mental model can be of how I want things to work.
It feels like that switch to test driven development where you start from the expected result and worry about the details later.
To your point, you can blow through damn-near anything pretty quickly now. Now I actually find myself problem-solving for nearly 8 hours every day. My brain feels fried at the end of the day way more than it used to.
I used to be like: "well, this thing will take me at least half a day, it's already 16:00, so I'll better do something quiet to cooldown to the end of the day and tackle this issue tomorrow". I'll leave the office in an regular mood and take the night to get ready for tomorrow.
Now I'm like: "17:30? 30 minutes? I have time to tackle another issue today!" I'll leave the office exhausted and take the night to try and recover from the day I had.
I agree it can be energizing because you can offload the bullshit work to a robot. For example, build me a CRUD app with a bootstrap frontend. Highly useful stuff especially if this isn't your professional forte.
The problems come afterwards:
1. The bigger the base codebase generation the less likely you're going to find time or energy to refactor LLM slop into something maintainable. I've spent a lot of time tailoring prompts for this type of generation and still can't get the code to be as precise as something an engineer would write.
2. Using an unfamiliar language means you're relying entirely on the LLM to determine what is safe. Suppose you wish to generate a project in C++. An LLM will happily do it. But will it be up to a standard that is maintainable and safe? Probably not. The devil is in the mundane details you don't understand.
In the case of (2) it's likely more instructive to have the LLM make you do the leg work, and then it can suggest simple verifiable changes. In the case of (1) I think it's just an extension of the complexity of any project professional or not. It's often better to write it correct the first time than write it fast and loose and then find the time to fix it later.
Red flag. In other words you don’t understand the implementation well enough to know if the AI has done a good job. So the work you have committed may work or it may have subtle artefacts/bugs that you’re not aware of, because doing the job properly isn’t of interest to you.
This is ‘phoning it in’, not professional software engineering.
At least when the AI does it you can review it.
Red flag again! If your protection is to "understand the implementation" it means buggy code. What makes a code worthy of trust is passing tests, well designed tests that cover the angles. LGTM is vibe testing
I go as far as saying it does not matter if code was written by a human who understands or not, what matters is how well it is tested. Vibe testing is the problem, not vibe coding.
But we're seeing that this becomes OK in the workplace, and I don't believe it is.
If you propose these changes that would've normally taken you 2 weeks as your own in a PR, then I, as the reviewer, don't know where your knowledge ends and the AI's hallucinations begin.
Do you need to do all of these things? Or is it because the most commonly forked template of this piece of code has this in its boilerplate? I don't know. Do you?
How can you make sure the code works in all situations if you aren't even familiar with the language, let alone the framework / API and protocol?
* Do you know that in Java you have to do string.Equals instead of == for equality?
* Do you know in Python that you assigning a new value to a function default persists beyond the function?
* And in JavaScript it does not?
* Do you know that the C# && does not translate to VB.NET's And?This has always been a problem in software engineering; of course -- sometimes staff have left, so you have to dig through tickets, related commits and documentation to intuit intent. But I think it's going to make for very weird drags on productivity in _new_ code that may not counter the acceleration LLMs provide, but will certainly exist.
vs. what this author is doing, which seems more like agent assisted coding than "vibe" coding.
With regard to the subject matter, it of course makes sense that managing more features than you used to be able to manage without $AI_MODEL would result in some mental fatigue. I also believe this gets worse the older you get. I've seen this within my own career, just from times of being understaffed and overworked, AI or not.
Then lots of people were introduced to the term "vibe coding" in these conversations, and so naturally took it as a synonym for using LLMs for coding assistance even when reading the code and writing tests and such.
Also because vibe coding just sounds cool.
A negative but courteous remark is "slamming", a tweet is an "attack", etc.
So yeah I'm not surprised that people conflate any use of AI with vibe-coding.
Personally I think "vibe-coding" has semantically shifted to mean any AI-assisted coding and we should just run with it. For the original meaning of vibe-coding, I suggest YOLO-Coding.
Agreed. I've seen some folks say that it requires absolute ignorance of the code being generated to be considered "vibe coded". Though i don't agree with that.
For me it's more nuanced. I consider a lack of review to be "vibed" related to how little you looked at it. Considering LLMs can do some crazy things, even a few ignored LOC might end up with a pretty "vibe coded" feelings, despite being mostly reviewed outside of those ignored lines.
Or here: https://en.wikipedia.org/wiki/Vibe_coding
Not looking at the code at all by default is essential to the term.
My interpretation is that you can look at the code but vibe coding means ultimately you're not writing the code, you're just prompting. It would make sense to prompt "I'd like variable name 'bar' to be 'foo' instead." and that would still be vibe coding.
The key difference is still the prompts and knowing what to reference/include in the context.
Do you believe that atrophy is not a real thing?
I've found that LLMs massively over-engineer things, regardless of the prompt used. How do you counter that without going back and forth at least a few times?
I've found that if an LLM writes too much code, even if I specified what it should be doing, I still have to do a lot of validation myself that would have been done while writing the code by hand. This turns the process from "generative" (haha) to "processing", which I struggle a lot more with.
Unfortunately, the reason I have to do so much processing on vibe code or large generated chunks of code is simply because it doesn't work. There is almost always an issue that is either immediately obvious, like the code not working, or becomes obvious later, like poorly structured code that the LLM then jams into future code generation, creating a house of cards that easily falls apart.
Many people will tell me that I'm not using the right model or tools or whatever but it's clear to me that the problem is that AI doesn't have any vision of where your code will need to organically head towards. It's great for one shots and rewrites, but it always always always chokes on larger/complicated projects, ESPECIALLY ones that are not written in common languages (like JavaScript) or common packages/patterns eventually, and then I have to go spelunking to find why things aren't working or why it can't generate code to do something I know is possible. It's almost always because the input for new code is my ask AND the poorly structured code, so the LLM will rarely clean up it's own crap as it goes. If anything, it keeps writing shoddy wrapper around shoddy wrappers.
Anyways, still helpful for writing boilerplate and segments of code, but I like to know what is happening and have control over how my code is structured. I can't trust the LLMs right now.
I could see other elements of isolation being useful, but this kind of feels like a lot of extra work and complexity which is part of the issue...
I take breaks.
But I also get drawn to overworking ( as I'm doing right now ), which I justify because "I'm just keeping an eye on the agent".
It's hard work.
It's hard to explain what's hard about it.
Watching as a machine does in an hour what would take me a week.
But also watching to stop the machine spin around doing nothing for ages because it's got itself in a mess.
Watching for when it gets lazy, and starts writing injectable SQL.
Watching for when it gets lazy, and tries to pull in packages it had no right to.
We've built a motor that can generate 1,000 horse power.
But one man could steer a horse.
The motor right now doesn't have the appropriate steering apparatus.
I feel like I'm chasing it around trying to keep it pointed forward.
It's still astronomically productive.
To abandon it would be a waste.
But it's so tiring.
The other tax is the intermittent downtime when you are waiting for the LLM to finish. In the olden days you might have productive downtime waiting for code to compile or a test suite to run. While this was happening you might review your assumptions or check your changes or realize you forgot an edge case and start working on a patch immediately.
When an LLM is running, you can't do this. Your changes are being done on your behalf. You don't know how long the LLM will take, or how you might rephrase your prompt if it does the wrong thing until you see and review the output. At best, you can context switch to some other problem but then 30 seconds later you come back into "review mode" and have to think architecturally about the changes made then "prompt mode" to determine how to proceed.
When you are doing basic stuff all of this is ok, but when you are trying to structure a large project or deal with multiple competing concerns you quickly overwhelm your ability to think clearly because you are thinking deeply about things while getting interrupted by completed LLM tasks or context switching.
Like when I'm asking it to run a bunch of tests against the UI using a browser tool, and something doesn't work. Then it goes and just writes code to update the database instead of using the user element.
My other thing that makes me insane is when I tell it what to do, and it says, "But wait, let me do something else instead."
Make it stop. Tell it to review whether the code is cohesive. Tell it to review it for security issues. Tell it to review it for common problems you've seen in just your codebase.
Tell it to write a todo list for everything it finds, and tell it fix it.
And only review the code once it's worked through a checklist of its own reviews.
We wouldn't waste time reviewing a first draft from another developer if they hadn't bothered looking over it and test it properly, so why would we do that for an AI agent that is far cheaper.
This statement resonates with me. Vibe coding gets the job done quickly, but without the same joy. I used to think that it was the finished product that I liked to create, but maybe it's the creative process of building. It's like LEGO kits, the fun is putting them together, not looking at the finished model.
On the flip side, coding sessions where I bang my head against the wall trying to figure out some black box were never enjoyable. Nor was writing POCOs, boilerplate, etc.
They ask a business question to the AI and it generates a bunch of code.
But honestly, coding isn't the part that slowed me down. Mapping the business requirements to code that doesn't fail is the hard part.
And the generated PRs are just answers to the narrow business questions. Now I need to spend time in walking it all back, and try to figure out what the actual business question is, and the overall impact. From experience I get very little answer to those questions.
And this is where Software Engineering experience becomes important. It's asking the right questions. Not just writing code.
Next to that I'm seeing developers drinking the cool-aid and submitting PRs where a whole bunch of changes are made, but they don't know why. Well, those changes DO have impact. Keeping it because the AI suggested it isn't the right answer. Keeping it because you agree with the AI's reasoning isn't the right answer either.
It's now 11:47am and I am mentally exhausted. I feel like my dog after she spends an hour at her sniff-training class (it wipes her out for the rest of the day.)
I've felt like that on days without the meetings too. Keeping up with AI tools requires a great deal of mental effort.
To people with little to no practical software experience, I can see why that seems incredible. Think of the savings! But to anyone who's worked in a legacy code base, even well written ones, should know the pain. This is worse. That legacy code base was at least written with intention, and is hopefully battle tested to some degree by the time you look at it. This is 20k lines of code written by an intern that you are now responsible for going through line by line, which is going to take at least as long as it would have to write yourself.
There are obvious wins from AI, and agents, but this type of development is a bad idea. Iteration loops need to be kept much smaller, and you should still be testing as you go like you would when writing everything yourself. Otherwise it's going to turn into an absolute nightmare fast.
...almost as if it's too eager to make its first commit. Much like a junior engineer might be.
It's not eager enough to iterate. Moreover, when it does iterate, it often brings along the same wrong solutions it came up with before.
It's way easier to keep an eye on small changes while iterating with AI than it is with letting it run free in a green field.
Even using it to spitball ideas can be a problem. I was using Claude to bounce ideas off of for a problem I was working on it, and it was dead set a specific solution involving a stack and some complex control logic was correct, when it reality it would have made the entire solution far more complicated. All I really needed was a sliding window into an array.
Usually that requires saying something, seeing if the other person understands what I'm saying, and occasionally repeating myself in a different way.
It can be real tiring when I'm with friends who only speak the other language so we're both using translator tools and basically repeating that loop up to 2-3 hours.
I've found the same situation with vibe coding. Especially when the model misunderstands what I want or starts going off on a tangent. sometimes it's easier to edit the original query or an earlier step in the flow and re-write it for a better result.
Choosing, analysing, verifying, talking to others rather than thinking in a sequential and organised way.
One of the reasons I moved away of that position was this constant fatigue by the end of the day. I was happy on seeing things moving forward but it felt that it took more energy to do less
If you're able to blaze through feature tickets using GenAI on existing projects of any major complexity, there's almost certainly something which would produce better code that you're skipping.
I have plenty of info and agents for Claude Code to take into account when I use it to make features in our projects, but what it can't know is the cadence of what our business partners expect, the unknown unknowns, conversations that humans have about the projects, and the way end-users feel about the project. My job is to direct it with those factors in mind, and the work to account for those factors takes time.
This fine for WFH/remote work. It didn't have great optics when I went back to in-office for a bit.
Maybe the fatigue comes from that mismatch?
The classical vibe coder style is to just ignore verification. That's not a good approach as well.
I think this space has not matured yet. We have old tools (test, lint) and some unreliable tools (agent assisted reviews), but nothing to match the speed of generation yet.
I do it by creating ad-hoc deterministic verifiers. Sometimes they'll last just a couple of PRs. It's cheap to do them now. But also, there must be a better way.
Coolest bit of research I cam across was what the brain does during sleep. It basically reduces connection during this. But it also makes you hallucinate (sleep). This was found in researching fish and also training LLMs there's great value in "forgetting" for generalization.
After studying it in LLMs for awhile I also came to your same conclusion about my own brain. Problems are often so complex you must let your brain forget in order to handle the complexity in the same sense I also believe this is the path to AGI.
However, when it comes to my professional work on a mature, advanced project, I find it much easier to write the code myself than to provide a very precise specification without which the LLM wouldn't generate code of a sufficiently high quality.