Part of me wishes it wasn't true but: AI coding is legit
I stay current on tech for both personal and professional reasons but I also really hate hype. As a result I've been skeptical of AI claims throughout the historic hype cycle we're currently in. Note that I'm using AI here as shorthand for frontier LLMs.
So I'm sort of a late adopter when it comes to LLMs. At each new generation of models I've spent enough time playing with them to feel like I understand where the technology is and can speak about its viability for different applications. But I haven't really incorporated it into my own work/life in any serious way.
That changed recently when I decided to lean all the way in to agent assisted coding for a project after getting some impressive boilerplate out of one of the leading models (I don't remember which one). That AI can do a competent job on basic coding tasks like writing boilerplate code is nothing new, and that wasn't the part that impressed me. What impressed me was the process, especially the degree to which it modified its behavior in practical ways based on feedback. In previous tests it was a lot harder to get the model to go against patterns that featured heavily in the training data, and then get it to stay true to the new patterns for the rest of the session. That's not true anymore.
Long story short, add me to the long list of people whose minds have been blown by coding agents. You can find plenty of articles and posts about what that process looks like so I won't rehash all the details. I'll only say that the comparisons to having your own dedicated junior or intern who is at once highly educated and dumb are apt. Maybe an even better comparison would be to having a team of tireless, emotionless, junior developers willing to respond to your requests at warp speed 24/7 for the price of 1/100th of one developer. You need the team comparison to capture the speed.
You've probably read, or experienced, that AI is good at basic tasks, boilerplate, writing tests, finding bugs and so on. And that it gets progressively worse as things get more complicated and the LoCs start to stack up. That's all true but one part that has changed, in more recent models, is the definition of "basic".
The bit that's difficult to articulate, and I think leads to the "having a nearly free assistant" comparisons, is what it feels like to have AI as a coding companion. I'm not going to try to capture it here, I'll just say it's remarkable.
The usual caveats apply, if you rely on agents to do extensive coding, or handle complex problems, you'll end up regretting it unless you go over every line with a magnifying glass. They will cheerfully introduce subtle bugs that are hard to catch and harder to fix when you finally do stumble across them. And that's assuming they can do the thing you're asking then to do at all. Beyond the basics they still abjectly fail a lot of the time. They'll write humorously bad code, they'll break unrelated code for no apparent reason, they'll freak out and get stuck in loops (that one suprised me in 2025). We're still a long way from agents that can actually write software on their own, despite the hype.
But wow, it's liberating to have an assistant that can do 100's of basic tasks you'd rather not be distracted by, answer questions accurately and knowledgeably, scan and report clearly about code, find bugs you might have missed and otherwise soften the edges of countless engineering pain points. And brainstorming! A pseudo-intelligent partner with an incomprehensibly wide knowledge base and unparalled pattern matching abilities is guaranteed to surface things you wouldn't have considered.
AI coding agents are no joke.
I still agree with the perspectives of many skeptics. Execs and middle managers are still out of their minds when they convince themselves that they can fire 90% of their teams and just have a few seniors do all the work with AI. I will read gleefully about the failures of that strategy over the coming months and years. The failure of their short sightedness and the cost to their organizations won't make up for the human cost of their decisions, but at least there will be consequences.
When it comes to AI in general I have all the mixed feelings. As an artist, I feel the weight of what AI is doing, and will do, to creative work. As a human I'm concerned about AI becoming another tool to funnel ever more wealth to the top. I'm concerned about it ruining the livelihoods of huge swaths of people living in places where there aren't systems that can handle the load of taking care of them. Or aren't even really designed to try. There are a lot of legitimate dystopian outcomes to be worried about.
Despite all that, actually using the technology is pretty exciting, which is the ultimate point of this post: What's your experience? Are you using agents for coding in practical ways? What works and what doesn't? What's your setup? What does it feel like? What do you love/hate about it?
LLMs are great for rubber ducking, helping you figure out what to do next, and writing generic boilerplate code that's been in its training data a million times (which is like half of your tasks in web development FWIW).
I genuinely don't believe it's good at anything else. I keep seeing posts like this that are shocked and amazed at how wonderful and efficient it is and I can't help but think of the study that found it actually causes more problems than it solved, despite people using it believing otherwise: https://arstechnica.com/ai/2025/07/study-finds-ai-tools-made-open-source-software-developers-19-percent-slower/
We're 3 years into the LLM craze and I've yet to see where the benefits are. If they were so good, why aren't they being used to easily fix open source issues? Even if it's just AI assisted? Because almost any code touched by AI that isn't ultra-generic is full of hard to find bugs and so many subtle issues it'll make the person reviewing your PR rather block you than deal with time wasting hallucinations.
There are like a million FOSS apps and tools out there and every time AI has a hand in "helping", it's obvious even when not mentioned. Because it produces bad code that looks right. And there hasn't been any major case of it actually helping fix major issues, the only thing I can see is that a lot of companies boasting using AI all the time coincidentally had the quality of their code/apps go down. See: Microsoft bragging about 30% of their code being written by AI, and suddenly new Windows 11 updates having weird and crazy bugs (like the task manager nonsense).
I mean, if your criteria when looking for AI code is “bad code that looks right”, definitionally you’ll only find bad code.
This feels like a CGI case, where all the CGI people remember from movies is bad… because they don’t notice the good examples to begin with, since the intention for most CGI is to be invisible.
The break in this reasoning is that we have strong CGI advocates. We have people who believe that CGI is that future and who would hold up as shining beacons any good examples. If AI had the capability of Silicon Valley's claims, they'd be right to. It would be revolutionary. It would be obvious, by looking at the rapidly growing collection of claims. We would have entire projects with thousands or millions of downstream users, maintained in whole or in large part by AI. Those projects are nowhere to be found. The only growing cries are from maintainers who are being overloaded in entirely new ways. The existence of bad CGI isn't the problem here, any tool can be misused, it's the lack of good CGI.
The study you mentioned comes up often in conversations about AI. I think it's valuable but limited. In part because of the small sample size, but more because of the conditions. There's a learning curve to figuring out what AI can do and what it can't. There's an even bigger learning curve in figuring out how to make it work effectively through MCP, rules, skills and other automatic prompt additions. If the study groups consisted of engineers instructed not to use AI versus engineers using AI who already had extensive experience with it (and the .md files to go with it), I suspect the results would be different. You can absolutely waste a lot of time trying to get AI to do everything for you, and therefore things it's just not good at, but once you know what not to use it for, and have built guardrails against its most common mistakes, the experience changes.
Do we know they aren't? In a practical workflow that uses AI assistance but isn't using AI to actually write all the code, I'm not sure you'd be able to tell that AI was involved at all.
I don't disagree with you that AI produces bad code that looks right, or that it isn't as great as some people have been saying it is. That's been my experience as well. But it's also been my experience that, used thoughtfully, it's incredibly helpful.
Looking forward it's hard to imagine any future that doesn't involve AI as an integral part of software development. But also there will be carnage along the way. Those subtle bugs we've been talking about are silently building up in codebases everywhere and that will only get worse. Not to mention unnecessary, difficult to maintain code that doesn't technically contain bugs. I chuckle when I hear people talk about how many lines of code they've written in last X days with AI that would have taken them weeks or months otherwise. X lines that should have actually been X/4 lines. Get back to us in 6 months when you're wading through that mess trying to figure out how to maintain it without scrapping it completely.
Caveat to that though: Thought it will no doubt plateau at some point, right now models are rapidly getting better with each iteration, their capabilities will most likely be dramatically better in a year.
It just depends on whether people keep in mind that it's just a tool. At work I see all the time people following it blindly like the answers it gives are gospel. It's like the issues we used to have with stack overflow on steroids. Imagine someone wholesale copying something they found on stack overflow proposing it as their own and then just copying the original comment from stack overflow as their description of the change. That effectively is what I see at work all the time and my leadership encourages usage of the tooling like this. They don't understand what they're talking about because nobody in leadership is going to have any experience with AI at all since it's so relatively new. They're all full of shit and I just have to deal with this and try to hold the fort down so the building doesn't burn down.
I've seen this at my work too, copilot can do a pretty good job at a lot of things but sometimes produces some garbage output, which my coworkers happily regurgitate. Just because copilot says something works some way, does not actually mean that's correct.
Coding with AI is easy and perfectly valid. We've had tools to check for program correctness for decades, linters, type checkers, unit tests etc.
Who cares if the "stohastic parrot" "hallucinates", when what it produces passes all possible tests and does what I asked? =)
Very apt. The models will happily add dependencies willy-nilly. I am constantly telling it to make a seam and separate the concerns into smaller classes.
"Don't add a dependency on NPC to our Sensor Service just to get the range of the currently active weapon/ability. Make an IRangeProvider and pass that in." If I could knock it upside the head, I would. It's good about doing those minor corrections, though, which makes it useful.
I have a little armchair-psychologist theory that using language models tickles the same pathways that addictive activities such as gambling or video games do. There's anticipation of the models next response, the novelty of the response, near-misses as the model gets it so close but not quite right, and variable reinforcement that all of that brings together. I'm not suggesting LLMs are addicting, just that I've felt that my use of it tends to feel similarly to how I feel when playing gambling type games (slots, roguelikes). The other week I caught myself eagerly anticipating using a language model to solve some problem, rather than eagerly anticipating the solution to the problem itself. I'm still a wee bit wary about the endgame of these tools.
That's valid, and I'd say they are addicting. There's a lot to be wary of. Always true with new tech, but maybe never more true than it is now.
And yeah they love dependencies, which is a symptom of how much their training data (people) love them.
The most profound article I've read on the topic, which supports my own findings, is that AI is just a tech-debt machine.
Yes, it will give you a functional prototype. It will get the bare-minimum to work, without considering readibility, maintainability or upgradability. You won't understand it, and anyone you pay to understand it, will prefer to write it their own way.
It writes what looks like working code, and is great for brainstorming, but it won't replace an engineer.
I haven't given it a serious try in a couple of years. Are there good tutorials now?
I haven't really looked. I have gotten some insights from various blog posts but sadly I don't have links.
My best tip is go slow and note everything it does that you don't want so that you can come up with rules to make it less of an idiot in future sessions. In the current generation of models there are consistent patterns of bad behavior that you have to account for. And be skeptical when it feels like magic, sometimes it really kind of is, but just as often it's an illusion.
My main complaint is that they can be more trouble than they are worth. But it’s up to me to try to build a heuristic for when that is the case. I’m not sure it’s possible to build such a heuristic that is anything close to optimal. So I lean towards using it less. That way I learn more anyway.
I've been integrating claude code into my workflow for the last month or so and for actually writing code it is quite useful. I would say upwards of 50% improvement in that area, going up or down depending on what the task is.
Now, of course, as a professional software engineer, the reality is that a minority of my time is spent actually writing code. So it's definitely not a 50% improvement in my overall efficiency. But just the hours it has spent, if you were to bill them, would easily recoup the cost.
And I'd also say that it's also just made me better at the rest of my job, because the parts of writing code that claude code is good at, is also some of the most boring and tedious and soul-sucking parts.
Some people try to make analogies about how it's like a junior engineer or whatever, but ultimately I don't think those kind of categorizations are useful. It is what it is - not quite like any kind of human. You just have to experiment to find good ways to fit in your workflow. There's no other way to figure out things that it'll be good and things it's bad at than just trying them.
I know I'm not a good enough coder to lean too heavily on code that it generates. If I can't understand what it's doing, I can ask it to explain it; if I don't get it, or am too lazy to learn, then I don't use the code. It's exceptionally stupid, but oh so very patient and so very fast to come up with more idiotic broken code. It's handy if I already know how to write this and I can easily fix the obvious mistakes.
What I do find is handy is using it essentially as a search engine: what tools are other people using for this type of problem? Send me a link of when folks are discussing this problem. What are some novel ways people are using this tool for? Then I read the human discussion on this, or download the human discussed useful tool.
The other use I'm getting sloppy on: explain X to me, summarize Y, find me the relevant pdf amongst these 50 where topic Z is discussed.
Insane value multiplier. Not even up for debate how useful and powerful it is.
Of course it's up for debate. It's an ongoing debate everywhere right now. The studies aren't in, the polls have barely been sent out, the conclusion is not done and written in stone.
I made an argument. You made a counter argument. Good interaction on the Internet!
I've long appreciated when someone has the ability to make a firm statement, designed to be arguable. Sometimes one must take a stance in order to illicit conversation.
You seem to not think it is useful or powerful, as I think it is.
Is the statement "This is not arguable" an argument? No facts or statements were put forth as proof or evidence.
I think it's a tool. It has uses, and I make use of it. But like the 10x engineer myth before it, I don't see any indications that multipliers are a thing that exist outside of games. The studies that have been published, while preliminary, hint at AI reducing productivity in the short term, and having the effect of loss of skill in the not-as-long-as-you-might-think term.