hckrnws

Super interesting study. One curious thing I've noticed is that coding agents tend to increase the code complexity of a project, but simultaneously massively reduce the cost of that code complexity.

If a module becomes unsustainably complex, I can ask Claude questions about it, have it write tests and scripts that empirically demonstrate the code's behavior, and worse comes to worst, rip out that code entirely and replace it with something better in a fraction of the time it used to take.

That's not to say complexity isn't bad anymore—the paper's findings on diminishing returns on velocity seem well-grounded and plausible. But while the newest (post-Nov. 2025) models often make inadvisable design decisions, they rarely do things that are outright wrong or hallucinated anymore. That makes them much more useful for cleaning up old messes.

Bad code has real world consequences. Its not limited to having to rewrite it. The cost might also include sanctions, lost users, attrition, and other negative consequences you don’t just measure in dev hours

Right, but that cost is also incurred by human-written code that happens to have bugs.

In theory experienced humans introduce less bugs. That sounds reasonable and believable, but anyone who's ever been paid to write software knows that finding reliable humans is not an easy task unless you're at a large established company.

The question then becomes, can LLMs generate code close to the same quality as professionals.

In my experience, they are not even close.

Well, if you keep in mind that "professionals" means "people paid to write code" then LLMs have been generating code at the same quality OR BETTER for about a year now. Most code sucks.

If you compare it to beautiful code written by true experts, then obviously not, but that kind of code isn't what makes the world go 'round.

We should qualify that kind of statement, as it’s valuable to define just what percentile of “professional developers” the quality falls into. It will likely never replace p90 developers for example, but it’s better than somewhere between there and p10. Arbitrary numbers for examples.

Can you quantify the quality of a p90 or p10 developer?

I would frame it differently. There are developers successfully shipping product X. Those developer are, on average, as skilled as necessary to work on project X. else they would have moved on or the project would have failed.

Can LLMs produce the same level of quality as project X developers? The only projects I know of where this is true are toy and hobby projects.

> Can you quantify the quality of a p90 or p10 developer?

Of course not, you have switched “quality” in this statement to modify the developer instead of their work. Regarding the work, each project, as you agree with me on from your reply, has an average quality for its code. Some developers bring that down on the whole, others bring it up. An LLM would have a place somewhere on that spectrum.

There was a recent study posted here that showed AI introduces regressions at an alarming rate, all but one above 50%, which indicates they spend a lot of time fixing their own mistakes. You've probably seen them doing this kind of thing, making one change that breaks another, going and adjusting that thing, not realizing that's making things worse.

The study is likely "SWE-CI: Evaluating Agent Capabilities in Maintaining Codebases via Continuous Integration". Regression rate plot is figure 6.

Read the study to understand what it is measuring and how it was measured. As I understand parent's summary is fine, but you want to understand it first before repeating it to others.

https://arxiv.org/abs/2603.03823

Observation 3

Bentley Software is proof that you can ship products with massive, embarrassing defects and never lose a customer. I can’t explain enterprise software procurement, but I can guarantee you product quality is not part of that equation.

This only helps if you notice the code is bad. Especially in overlay complex code, you have to really be paying attention to notice when a subtle invariant is broken, edge case missed, etc.

Its the same reason a junior + senior engineer is about as fast as a senior + 100 junior engineers. The senior's review time becomes the bottleneck and does not scale.

And even with the latest models and tooling, the quality of the code is below what I expect from a junior. But you sure can get it fast.

This is the most important point in the thread. The study measures code complexity but the REAL bottleneck is cognitive load (and drain) on the reviewer.

I've been doing 10-12 hour days paired with Claude for months. The velocity gains are absolutely real, I am shipping things I would have never attempted solo before AI and shipping them faster then ever. BUT the cognitive cost of reviewing AI output is significantly higher than reviewing human code. It's verbose, plausible-looking, and wrong in ways that require sustained deep attention to catch.

The study found "transient velocity increase" followed by "persistent complexity increase." That matches exactly. The speed feels incredible at first, then the review burden compounds and you're spending more time verifying than you saved generating.

The fix isn't "apply traditional methods" — it's recognizing that AI shifts the bottleneck from production to verification, and that verification under sustained cognitive load degrades in ways nobody's measuring yet. I think I've found some fixes to help me personally with this and for me velocity is still high, but only time will tell if this remains true for long.

The part that gets me is when it passes lint, passes tests, and the logic is technically correct, but it quietly changed how something gets called. Rename a parameter. Wrap a return value in a Promise that wasn't there before. Add some intermediate type nobody asked for. None of that shows up as a failure anywhere. You only notice three days later when some other piece of code that depended on the old shape breaks in a way that has nothing to do with the original change.

[dead]

> The study found "transient velocity increase" followed by "persistent complexity increase."

Companies facing this reality are of course typically going to use AI to help manage the increased complexity. But that leads quickly to AI becoming a crutch, without which even basic maintenance could pose an insurmountable challenge.

>The fix isn't "apply traditional methods"

I would argue they are. Those traditional methods aim at keeping complexity low so that reading code is easier and requires less effort, which accelerates code review.

[dead]

I find LLMs get much more prone to making mistakes or missing references when the size or complexity of the code increases. I have a “vibe coded” application that is just for personal use, and I’ll usually create a fresh prompt after a large refactor and ask “were all references to the previous approach removed, and has the application been fully migrated to using the new approach?”

It finds spots it missed during the refactor basically every time.

So I partially agree with you, but I think it takes multiple passes and at least enough understanding to challenge the LLM and ask pointed questions.

What happens if you or future developers become unable to access Claude, the proprietary product of Anthropic?

The open source models are pretty good too now. They are a few months behind, but not more than that. Sure, you still have to host them in the cloud to get enough VRAM to run them - but looking 10+ years into the future the end game here will probably be having a local LLM that runs on your own computer and is more than capable enough to do coding for you.

I've been asking myself the same question. Realistically I think it depends a lot on how many providers are available in the future. If you lose access to one you can move to another, its not a single point of failure per say. I think this question gets a lot more relevant if the providers get more monopolized instead of gaining wider spread, so far we've only seen more providers appear.

> empirically demonstrate the code's behavior

That is completely insufficient for code of any real complexity. All this does is replacing known bugs with unknown bugs.

I think thats a fallacy. As of right now there is a point of no return where the complexity cant be broken by the agent itself without breaking more on other things. I’ve seen it before. Agents cheat on tests, break lint and type rules.

I was hoping for it to work, but It didn’t for me.

Still trying to figure out how to balance it.

> have it write tests

Just make sure it hasn't mocked so many things that nothing is actually being tested. Which I've witnessed.

I’ve also seen Opus 4.5 and 4.6 churn out tons of essentially meaningless tests, including ones where it sets a field on a structure and then tests that the field was set.

You have to actually care about quality with these power saws or you end up with poorly-fitting cabinets and might even lose a thumb in the process.

The first thing you should do after having them write tests is delete half of the tests.

> but simultaneously massively reduce the cost of that code complexity.

Citation needed. Until proven otherwise complexity is still public enemy #1. Particularly given that system complexity almost always starts causing most of its problems once a project is further along I don’t think we will know anything meaningful about that statement for at least a year.

> Super interesting study. One curious thing I've noticed is that coding agents tend to increase the code complexity of a project, but simultaneously massively reduce the cost of that code complexity.

This is the same pattern I observed with IDEs. Autocomplete and being able to jump to a definition means spaghetti code can be successfully navigated so there's no "natural" barrier to writing spaghetti code.

Yes, it's not surprising that warnings and complexity increased at a higher rate when paired with increased velocity. Increased velocity == increased lines of code.

Does the study normalize velocity between the groups by adjusting the timeframes so that we could tell if complexity and warnings increased at a greater rate per line of code added in the AI group?

I suspect it would, since I've had to simplify AI generated code on several occasions but right now the study just seems to say that the larger a code base grows the more complex it gets which is obvious.

"Notably, increases in codebase size are a major determinant of increases in static analysis warnings and code complexity, and absorb most variance in the two outcome variables. However, even with strong controls for codebase size dynamics, the adoption of Cursor still has a significant effect on code complexity, leading to a 9% baseline increase on average compared to projects in similar dynamics but not using Cursor."

To add to the person who quotes the relevant part of the study, they also point that the velocity increase disappears after a month or two.

That was my thought as well, because obviously complexity increases when a project grows regardless of AI

Yeah, I have a more complex project I'm working on with Claude, but it's not that Claude is making it more complex; it's just that it's so complex I wouldn't attempt it without Claude.

There are actually quite a few studies out there that look at LLM code quality (e.g. https://scholar.google.com/scholar?hl=en&as_sdt=0%2C5&q=LLM+...) and they mostly have similar findings. This reinforces the idea that LLMs still require expert guidance. Note, some of these studies date back to 2023, which is eons ago in terms of LLM progress.

The conclusion of this paper aligns with the emerging understanding that AI is simply an amplifier of your existing quality assurance processes: Higher discipline results in higher velocity, lower discipline results in lower stability (e.g. https://dora.dev/research/2025/) Having strong feedback and validation loops is more critical than ever.

In this paper, for instance, they collected static analysis warnings using a local SonarQube server, which implies that it was not integrated into the projects they looked at. As such these warnings were not available to the agent. It's highly likely if these warnings were fed back into the agent it would fix them automatically.

Another interesting thing they mention in the conclusion: the metrics we use for humans may not apply to agents. My go-to example for this is code duplication (even though this study finds minimal increase in duplication) -- it may actually be better for agents to rewrite chunks of code from scratch rather than use a dependency whose code is not available forcing it to instead rely on natural language documentation, which may or may not be sufficient or even accurate. What is tech debt for humans may actually be a boon for agents.

Code duplication is normally referring to duplicate code within the same code base, not writing something yourself instead of using a library.

> We find that the adoption of Cursor leads to a statistically significant, large, but transient increase in project-level development velocity, along with a substantial and persistent increase in static analysis warnings and code complexity. Further panel generalized-method-of-moments estimation reveals that increases in static analysis warnings and code complexity are major factors driving long-term velocity slowdown. Our study identifies quality assurance as a major bottleneck for early Cursor adopters and calls for it to be a first-class citizen in the design of agentic AI coding tools and AI-driven workflows.

So overall seems like the pros and cons of "AI vibe coding" just cancel themselves out.

The part you quoted doesn't support your conclusion. Per your quoted paragraph, the benefit of "AI vibe coding" is a large, but transient (i.e temporary) increase in development velocity; while the drawback is a persistent increase in static analysis warnings and code complexity.

To me, this sounds like after the transient increase of velocity has died down, you're left with the same development velocity as you had when you started, but a significantly worse code base.

The implication seems to be that if quality assurance is prioritized, the negative impact would be eliminated.

This seems to assume the main cause is the accumulation of defects due to lack of static analysis and testing.

I think a more likely cause is, the code begins to rapidly grow beyond the maintainers' comprehension. I don't think there is a technical solution for that.

> This seems to assume the main cause is the accumulation of defects due to lack of static analysis and testing.

Neither (current) static analysis nor testing is sufficient to score the commit on complexity.

As a trivial example (i.e. probably not something the LLM would do), if the code was a series of 25 if-then-else statements when it could have been a lookup table, no tool is going to flag that.

Now imagine what patterns, which non-junior devs would reject without even thinking, that an LLM would inject. No test, nor any static analysis tool, is going to flag that non-essential complexity.

Another thing, tangential to what you talk about: if you have a junior programmer who does write some silly pattern, such as a 25 long if/else chain, you can notice that in code review and have a conversation with the junior programmer. Talk with them the possible alternatives; talk about switch, talk about lookup tables, discuss pros and cons of each solution. In the end, you have a better, slightly less junior programmer who will be more conscious of the choices they're making in the future. Repeat a few dozen times across a few years and you have a senior programmer.

You can't do this with language models. They don't learn. You can write in your CLAUDE.md stuff like "avoid long if/else chains" of course, but 1) cluttering the context with lots and lots of micro-managing instructions has its own downsides, and 2) you don't want a list of rules, you want deliberate weighing of the trade-offs involved in choices made.

The "transient velocity, persistent complexity" finding tracks with what I've seen. But I think the real issue is most people treat AI code like it's done when it runs. It's not. It's a first draft.

I've been using Claude pretty heavily for the last month or so on a few side projects. The speed is genuinely nuts — stuff that would've taken me a weekend gets roughed out in an afternoon. But if I just commit what it gives me, the codebase turns into this weird verbose mess where everything kinda works but nothing is clean.

What actually helped was treating AI output the same way you'd treat code from a fast but sloppy contractor: run it, write real tests (not the fluffy ones it generates that mock everything into oblivion), then refactor before committing. Basically the build-test-refactor-commit cycle that dalemhurley mentioned. The models are good enough to generate that first pass, but you still need taste to shape it.

The 9% complexity increase makes total sense if people skip that step. Which, let's be honest, most of us do when we're moving fast.

I really like AI and find it very useful, but it's certainly a good idea to keep it on a short leash because if you don't it can run away and get lost.

Really interesting study. One thing I keep coming back to is that tests have no way of catching this sort of tech debt. The agent can introduce something that will make you rip your hair out in 6 months, but tests are green...

My theory is that at least some of this is solvable with prompting / orchestration - the question is how to measure and improve that metric. i.e. how do we know which of Claude/Codex/Cursor/Whoever is going to produce the best, most maintainable code *in our codebase*? And how do we measure how that changes over time, with model/harness updates?

This matches what I've seen building with AI assistance, velocity goes up fast, but you start accumulating complexity debt you didn't consciously design. The difference is intentionality. When the architecture and systems decisions are yours and you're using AI to execute, the complexity stays manageable. When AI drives both the architecture AND the code, that's when this paper's findings kick in hard.

The study only looks at what lands in the PR. In my experience a single prompt can trigger 20+ tool calls, most of them reads and greps. The final diff is a tiny fraction of what actually happened. Hard to judge quality without seeing the process.

Interesting findings. I use AI agents (Claude, Windsurf) exclusively to build production software without being a developer myself. Speed is real but so is context drift – the AI breaks unrelated things while fixing others. Git became essential for me because of this.

They're measuring development speed through lines of code. To show that's true they'd need to first show that AI and humans use the same number of lines to solve the same problem. That hasn't been my experience at all. AI is incredibly verbose.

Then there's the question of if LoC is a reliable proxy for velocity at all? The common belief amongst developers is that it's not.

Was it Bill Gates who likened LoC to measuring airplane construction progress by weight?

See also

-2000 lines of code

https://news.ycombinator.com/item?id=26387179

This is actually one thing I have found LLMs surprisingly useful for.

I give them a code base which has one or two orders of magnitude of bloat, and ask them to strip it away iteratively. What I'm left with usually does the same thing.

At this point the code base becomes small enough to navigate and study. Then I use it for reference and build my own solution.

> They're measuring development speed through lines of code.

Yeah, this is the biggest facepalm.

Didn't we grow out of this idiocy 40 years ago? This shit again? Really?

This study's cutoff date was August 2025. I don't think this result is surprising given the level of coding agent ability back then. The whole thing just shows how out-of-date academic publishing is on this subject.

>This yields 806 repositories with adoption dates between January 2024 and March 2025 that are still available on GitHub at the time of data analysis (August 2025).

There were very few people who thought that coding agents worked very well back then. I was not one of them, but I _do_ think they work today.

Evergreen excuses for tech people desperately want to work. I get why, it would give you agency to do other things you WANT to do. I tried reviewing a colleagues agent-generated code and it was practically unreviewable. I watched him blame himself, saying he just needed to adjust a parameter. He tried everything except admit the machine does not conceptionaly understand what he was asking.

we're one more rl run from Codex not trying to satisfy the type checker by replacing an index accessor with a BFS of all the keys in the API response and matching the correct property via regex.

one more, i swear.

Eh, I don't know. I mean, are we seeing better models now? Of course. But are they truly leaps and bounds better? No, and I get confused by people saying that they are. They're better but not like... 10x better.

And when people were studying ChatGPT 3.5, everyone would go "Oh, but that wasn't 4!", and when people talk about Opus 4.5 they go "4.6 is so much better!".

My personal position right now is that people are extremely bad at evaluating model output/ changes in model capabilities. Model benchmarks do not reflect the position that models are just 10x better than they were a year ago, but with how people discuss them you'd think that 10x was underselling it.

I don't have personal experience, but there seems to be a broad consensus that Opus 4.5 was tipping point between "kinda bad" and "actually kinda useful".

So a cutoff point of August 2025 just before that is a bit unfortunate (I'm sure they'll be newer studies soon).

That reaction has happened with every model release for the past few years. Maybe they aren’t the same people, but it’s always “old model was terrible, new model gets it right” then “new model was terrible, newer model gets it right,” ad infinitum.

A large proportion of my professional network were in the "AI for code generatin might just be a fad" camp pre Opus 4.5 (and the Codex/Gemini models that came out shortly after that), and now almost everyone seems to think that AI will have at least some place in professional development environments on an ongoing basis.

I've recently given it a go myself, and it certainly doesn't get it right all the time. But I was able to generate AI-assisted code that met my quality standards at roughly the same speed as coding it by hand.

FWIW I am definitely someone who uses AI. I have been using it for a few years now. There's no question that models have improved. I'd say the biggest leap was around the ChatGPT 3.5 -> 4.0, which radically reduced hallucination problems. The big issue of "it just made up a module that doesn't exist" more or less went away at that point. This was the big leap from "spits out text that might help you" to "can produce value".

Since then it has been incremental. I would say the big win has been that models degrade more slowly as context grows. This means, especially for heavily vibecoded-from-scratch projects, that you hit the "I don't even know wtf this is anymore" wall way later, maybe never if you're steering things properly.

I think because you can avoid hitting that wall for longer, people see this as radically different. It's debatable whether that's true or not. But in terms of just what the model does, like how it responds to prompts, I genuinely think it is only marginally better. And again, I think benchmarks confirm this, and I quite like Fodor's analysis on benchmarking here[0].

I use these models daily and I try new models out. I think that people over emphasize "model did something different" or "it got it right" when they switch over to a new model as "this is radically better", which I believe is simply a result of cognitive bias / poor measurement.

[0] https://jamesfodor.com/2025/06/22/line-goes-up-large-languag...

I have experience and the gap is exaggerated imo. edit: And I think benchmarks largely support this, and benchmarks are already biased to overstate LLM performance IMO.

This is the perennial excuse, and I'm sure we'll continue to see it. Folks will say the exact same thing when the current crop of slop-generators have been replaced by a newer ilk

I think the issue is people AI assisted code, test then commit.

Traditional software dev would be build, test, refactor, commit.

Even the Clean Coder recommends starting with messy code then tidying it up.

We just need to apply traditional methods to AI assisted coding.

At that point you lose all potential gains from just using AI - it's harder and slower to read and understand a bunch of verbose slop and then clean it up

Depends on the nature of the tool I would imagine - eg. Claude Code Terminal (say) would have higher entry requirements in terms of engineering experience (Cursor was sold as newbie-friendly) so I would predict higher quality code than Cursor in a similar survey.

ofc that doesn't take into account the useful high-level and other advantages of IDEs that might mitigate against slop during review, but overall Cursor was a more natural fit for vibe-coders.

This is said without judgement - I was a cheerleader for Cursor early on until it became uncompetitive in value.

Now someone do a research study where a summary of this research paper is in the AGENTS.md and let’s see if the overall outcomes are better

[dead]

[dead]

[dead]

[dead]

[flagged]

[flagged]

I hope this is just an edit away from making more sense to me.

Interesting from an historical perspective. But data from 4/2025? Might as well have been last century.

I think the gist of it still applies to even Claude Code w/Opus 4.6.

It's basically outsourcing to mediocre programmers - albeit very fast ones with near-infinite patience and little to no ego.

In my experience, 4.6, together with the Claude Code improvements was a non-linear event. A threshold was crossed that forced me to review my complete model of genAI.

If your ideas/studies/experiences with genAI for software development and engineering were from before januari, basically /clear and re-init.

It doesn't map well to a mediocre human programmer, I think. It operates in a much more jagged world between superhuman, and inhuman stupidity.

A data center of geniuses on a medium dose of LSD.

AI is not perfect sure, one has to know how to use it. But this study is already flawed since models improved a lot since the beginning of 2026.

This is not a useful, constructive or meaningful statement.

Attempting to claim the models are the future by perpetually arguing their limitations are because people are using the models wrong or that the argument has been invalidated because the new model fixes it might as well be part of the training data since Claude Opus 3.5.

No no, I didn't say that at all. I'm just saying that the studies are irrelevant since models got a boost in their competence. I'm not in a fight pro or against llm's, I know how they work and their limitations. But the complexity of the problems they solve increased since opus 4.5 . If you can't admit that, it's your problem.

Also, I'm not blaming users for their shortcomings. I'm just saying they are not perfect but you can get different outcomes according to how you use them.

> Also, I'm not blaming users for their shortcomings. I'm just saying they are not perfect but you can get different outcomes according to how you use them.

I don’t question your sincerity. The problem is these arguments and framing are the same ones used by AI hucksters because it’s trivial to move the goalposts. They pull the same tricks as an MLM: success is just around the corner, forever.

IMHO, there’s a chasm between how to actually use these models and what they are being sold as.

[dead]