In my experience, it’s even more effort to get good code with an agent-when writing by hand, I fully understand the rationale for each line I write. With ai, I have to assess every clause and think about why it’s there. Even when code reviewing juniors, there’s a level of trust that they had a reason for including each line (assuming they’re not using ai too for a moment); that’s not at all my experience with Codex.

Last month I did the majority of my work through an agent, and while I did review its work, I’m now finding edge cases and bugs of the kind that I’d never have expected a human to introduce. Obviously it’s on me to better review its output, but the perceived gains of just throwing a quick bug ticket at the ai quickly disappear when you want to have a scalable project.

I was careful to say "Good code still has a cost" and "delivering good code remains significantly more expensive than [free]" rather than the more aesthetically pleasing "Good code is expensive.

I chose this words because I don't think good code is nearly as expensive with coding agents as it was without them.

You still have to actively work to get good code, but it takes so much less time when you have a coding agent who can do the fine-grained edits on your behalf.

I firmly believe that agentic engineering should produce better code. If you are moving faster but getting worse results it's worth stopping and examining if there are processes you could fix.

Code is cheaper. Simple code is cheap. More complex code may not be cheaper.

The reason you pay attention to details is because complexity compounds and the cheapest cleanup is when you write something, not when it breaks.

This last part is still not fully fleshed out.

For now. Is there any reason to not expect things to improve further?

Regardless, a lot of code is cheap now and building products is fun regardless, but I doubt this will translate into more than very short-term benefits. When you lower the bar you get 10x more stuff, 10x more noise, etc. You lower it more you get 100x and so on.

Computer programming is cheap. Software engineering is expensive.

I think the cost and work remains the same. What has change is efficiency. Previously people had to manually program byte after byte. Then came C and streamlined it, allowing faster development.

With python I can write a simple debugging UI server with a few lines.

There are frameworks that allow me to complete certain tasks in hours.

You do not need to program everything from scratch.

The more code, the faster everything gets, since the job is mostly done.

We are accelerating, but we still work 9 to 5 jobs.

Definitely the market incentives for "good code" have never been worse, but I'm wouldn't be so sure the cost of migrating decent pieces of generated code to good code is worse than writing good code from whole cloth.

Spaghetti code was always a thing though

All the hype is on how fast it is to produce code. But the actual bottleneck has always been the cost of specifying intent clearly enough that the result is changeable, testable, and correct AND that you build something that brings value.

I'd add in "code is easier to write than it is to read" - hence abstraction layers designed to present us with higher level code, hiding the complex implementations.

But LLMs are both really good at writing code _and_ reading code. However, they're not great at knowing when to stop - either finishing early and leaving stuff broken, over-engineering and adding in stuff that's not needed or deciding it's too hard and just removing stuff it deems unimportant.

I've found a TDD approach (with not just unit tests but high-level end-to-end behaviour-driven tests) works really well with them. I give them a high-level feature specification (remember Gherkin specifications?) and tell it to make that pass (with unit tests for any intermediate code it writes), make sure it hasn't broken anything (by running the other high-level tests) then, finally, refactor. I've also just started telling it to generate screenshots for each step in the feature, so I can quickly evaluate the UI flow (inspired by Simon Willison's Rodney tool).

Now I don't actually need to care if the code is easy to read or easy to change - because the LLM handles the details. I just need to make sure that when it says "I have implemented Feature X" that the steps it has written for that feature actually do what is expected and the UI fits the user's needs.

> Do current LLM based agents generate code which is easy to change?

They do. I am no longer writing code, everything I commit is 100% generated using an agent.

And it produces code depending on the code already in my code-base and based on my instructions, which tell it about clean-code, good-practices.

If you don't get maintainable code from an LLM it's for this reason: Garbage in, garbage out.

> Once you can ask your agent to change a feature and be 100% sure they won't break other features then you don't care about how the code looks like.

That bar is unreasonably high.

Right now, if I ask a senior engineer to change a feature in a mature codebase, I only have perhaps 70% certainty they won't break other features. Tests help, but only so far.

I am constantly getting LLMs to change features and fix bugs. The key is to micromanage the LLM and its context, and read the changes. It's slower that vibe coding but faster than coding by hand, and it results in working, maintainable software.

We won't be able to be sure of 100% with LLMs but maybe proper engineering around evals get us to an acceptable level of quality based on the blast radius/safety profile.

I'd also argue that we should be pushing towards tracer bullets as a development concept and less so prototypes that are nice but meant to be thrown away and people might not do that.

The clean room auto porting, after a messy exploratory prototyping session would be a nice pattern, nonetheless.

Irrelevant because you are not going to make new changes by hand. You will use AI for that.

> The new skill is mastering the craft of directing cheap inputs toward valuable outcomes.

Strongly agree with this. It took me awhile to realize that "agentic engineering" wasn't about writing software it was about being able to very quickly iterate on bespoke tools for solving a very specific problem you have.

However, as soon as you start unblocking yourself from the real problem you want to solve, the agentic engineering part is no longer interesting. It's great to be solving a problem and then realize you could improve it very quickly with a quick request to an agent, but you should largely be focused on solving the problem.

Yet I see so many people talking about running multiple agents and just building something without much effort spent using that thing, as though the agentic code itself is where the value lies. I suspect this is a hangover from decades where software was valuable (we still have plenty of highly valued, unprofitable software companies as a testament to this).

I'm reminded a bit of Alan Watts' famous quote in regards to psychedelics:

> If you get the message, hang up the phone.

If you're really leveraging AI to do something unique and potentially quite disruptive, very quickly the "AI" part should become fairly uninteresting and not the focus of your attention.

Or another way of looking at it: just because digging a ditch became cheap and fast with the backhoe doesn't mean you can just dig a bunch of ditches and become rich.

Indeed: The act of actually typing the code into an editor was never the hard or valuable part of software engineering. The value comes from being able to design applications that work well, with reasonable performance and security properties.

Raising $100M doesn’t even mean you have a good idea or an idea people like or an idea you can even make money on.

I think we’re falling into a trap of overestimating the value of incrementally directing it. The output is all coming from the same brain so what stops someone just getting lucky with a prompt and generation that one-shots the whole thing you spent time breaking down and thinking about. The code quality will be the same, and unless you’re directing it to the point where you may as well be coding the old way, the decision-making is the same too.

> I find that even though the cost of writing code is immensely cheap, reviewing and validating that it works in certain code bases (like the millions of line mono repo I work in at my job) is extremely high.

That is my observation as well. Churning code is easy, but making sure the code is not total crap is a completely new challenge and concern.

It's not like prior to LLMs code reviews didn't required work. Far from it. It's just that how the code is generated in a completely different way, and in some cases with barely any oversight from vibecoders who are trying to punch way above their weight. So they generate these massive volumes of changes that fail in obvious and subtle ways, and the flow is relentless.

For most non-hobby project, the cost of code was in breaking a working system (whether by a bona fide bug, or a change in some unspecified implicit assumption). That made changes to code incredibly expensive - often much more than the original implementation.

It sounds harsh, but over the lifetime of a project, 10-lines/person/day is often a high estimate of the number of lines produced. It’s not because humans type so slow - it is because after a while, it’s all about changing previously written lines in ways that don’t break things.

LLMs are much better at that than humans, if the constraints and tests are reasonably well specified.

It doesn't disappear but it does easy up some instances of fighting configuration, documentation, syntax or even comparing three approaches that are similar but you don't know their full effect.

I think it's a very fun space, finally being able to empower many people who in the past wouidve been bottlenecked unless they were using very simple tools for their domain and upskilled enough. Those 2 things will still be true, but the speed at which some things can happen at the exploration and other layers has seen a significant speedup.

Other problems like entropy/slop, security, system testing, lack of automation fundamentals arise but it's a good problem to start tackling.

I'm very focused on evals [1] because is what allows me to not to be the bottleneck with economists who I want to empower to code end to end and I'd like that mental shift to happen for anyone becoming a builder so non traditional developers and developers by trade have a common language for product building [2]. That part of speaking to different audiences and combating hype that promises to do everything for you Vs the intent that's actually needed is hard, but trying gets you to advance quite a lot.

[1] https://alexhans.github.io/posts/series/evals/measure-first-...

[2] https://ai-evals.io

> LLMs make the typing cheap, but they don’t make the reasoning cheap.

LLMs lower the cost of copy/pasting code around, or troubleshooting issues using standard error messages.

Instead of going through Stack Overflow to find how to use a framework to do some specific thing, you prompt a model. You don't even need to know a thing about the language you are using to leverage a feedback loop.

LLMs lower the cost of a multitude of drudge work in developing software, such as having to read the docs to learn how a framework should be used to achieve a goal. You still need to know what you are doing, but you don't need to reinvent the wheel.

The focus is on downstream, but is upstream ready for this speed up?

The linked blog post draws comparisons to the industrial revolution however in the industrial revolution the speed up caused innovation upstream not downstream.

The first innovation was mechanical weaving. The bottleneck was then yarn. This was automated so the bottleneck became cotton production, which was then mechanised.

So perhaps the real bottleneck of being able to write code faster is upstream.

Can requirements of what to build keep up with pace to deliver it?

Nice to see you here (Just reached out on bluesky over sandboxing - gandolin). I follow your work and agree and am hoping that you and others who have well earned audiences based on awesome open source work, can help with the advocacy on mental shifts, not just for developers but also non Devs that become builders.

I'm very focused on their minimalistic building experience as a way to make me and other traditional developers, not the bottleneck and empowering them end to end.

I think AI evals [1] are a big part of that route and hope that different disciplines can finally have probable product design stories [2] instead of there being big gaps of understanding between them.

[1] https://alexhans.github.io/posts/series/evals/measure-first-...

[2] https://ai-evals.io

I don’t think we can expect all workers at all companies to just adopt a new way of working. That’s not how competition works.

If agentic AI is a good idea and if it increases productivity we should expect to see some startup blowing everyone out of the water. I think we should be seeing it now if it makes you say ten times more productive. A lot of startups have had a year of agentic AI now to help them beat their competitors.

One of the most interesting aspects is when LLMs are cheap and small enough so that apps can ship with a builtin one so that it can adjust code for each user based on input/usage patterns.

>but medium and long term we need to figure out how to build systems in a way that it can keep up with this increased influx of code.

Why? Why do we need to "write code so much faster and quicker" to the point we saturate systems downstream? I understand that we can, but just because we can, does'nt mean we should.

I was having this conversation at work, where if the promise of AI coding becomes true and we see it in delivery speed, we would need to significantly increase the throughput of all other aspects of the business.

Totally agree - that's what I was trying to get at with "organizational habits". The way we plan, organize and deliver software projects is going to radically change.

I'm not ready to write about how radically though because I don't know myself!

> If we want to keep shipping at this speed

Do we? Spewing features like explosive diarrhea is not something I want.

The linked article is worth reading alongside this one.

The thing I'd add from running agents in actual production (not demos, but workflows executing unattended for weeks): the hard part isn't code volume or token cost. It's state continuity.

Agents hallucinate their own history. Past ~50-60 turns in a long-running loop, even with large context windows, they start underweighting earlier information and re-solving already-solved problems. File-based memory with explicit retrieval ends up being more reliable than in-context stuffing - less elegant but more predictable across longer runs.

Second hard part: failure isolation. If an agent workflow errors at step 7 of 12, you want to resume from step 6, not restart from zero. Most frameworks treat this as an afterthought. Checkpoint-and-resume with idempotent steps is dramatically more operationally stable.

Agree it's not just habits - the infrastructure mental model has to change too. You're not writing programs so much as engineering reliability scaffolding around code that gets regenerated anyway.

Writing code is cheaper than ever. Maintaining it is exactly the same as ever and it scales with the LOC.

Code is still liability but it's undeniable that going from thought to running code is very cheap today.

> We don't know what the actual cost is yet.

We kind of do? Local models (thought no state of the art) set a floor on this.

Even if prices are subsidized now (they are) that doesn't mean they will be more expensive later. e.g. if there's some bubble deflation then hardware, electricity, and talent could all get cheaper.

Are you familiar with the cathedral vs the bazaar?

Dollars to donuts that at some point someone is going to discover that senior engineers spend just as much time reviewing, fixing, and dealing with blowups caused by, shitty AI-generated code produced by more junior coders....as they did providing various forms of mentoring of said junior coders, except those junior coders become better developers in the latter case, whereas the AI generates the same shitty results or even worse, inconsistent quality code.

Yeah, it’s odd watching the outsourcing debate play out again. The results are gonna be the same.

Which is a shame, cause I think LLMs have a lot more use for software dev than writing code. And that’s really what’s going to shift the industry - not just the part willing to cut on quality.

I think the prototype thing is absolutely true but breaks down like all prototypes at the level of collaborating, sharing and evolving while handling entropy throug simplicity UNLESS you know what you're doing or the agent steers you with very opinionated tooling customized to your context. I'm thinking about empowering people to be builders and less so a software developer who can make the right tradeoffs.

Empowering people to work Tracer bullet style after they've selected their prototype of choice and thrown it away might be a powerful pattern that actually gets us into a nice collaborative space.

This feels to me like peak sfba mentality on par with "move fast and break things". Outside of trying to create a unicorn, is this really how people create things?

It seems to me that in order to obtain the ability to build things that other people like, you need to go through the process of creating things they won't. Like a painter needs to paint a bunch of crappy paintings to learn how to create a good painting. If you have the LLM create these throwaway prototypes, how will you even know when you come across a good idea and how will you be able to build it.

Possibly even more important than knowing where to hit it (what to code), is knowing where not to hit it (what not to code). Hitting the thing in the wrong place can lead to catastrophe. Making a code change you don't need can blow up production or paint your architecture into a corner.

AIs so far seem to prefer addition by addition, not addition by subtraction or addition by saying "are you sure?".

This doesn't mean that "code is cheap" is bad. Rather, it means that soon our primary role will be to guide AIs to produce a high proportion of "code that was cheap", while being able to quickly distinguish, prevent, and reject "cheap code".

Code you can’t just throw away is a liability because you have to keep supporting it / servicing it. Claude Code and friends also change that part of the cost equation:

You might not get gcc/llvm level optimization from a newly built compiler - but if you had a home-built one, which took $15,000/month engineer to support (for years!) you can now get a new one for $20,000 every 3 months, for a 50% cost saving, each time changing your requirements (which you couldn’t do before).

Code used to be a liability, like a car or an apartment for the average person. Now it’s a liability, like a car or apartment for Bill Gates.

I would normally agree, but I think the "code is a liability" quote assumes that humans are reading and modifying the code. If AI tools are also reading and modifying their own code, is that still true?

>Code is a liability, not an asset

Then "AI" code is even more of a liability.

I think you mean to say, "code you don't understand is a liability, not an asset"

But please correct me if I'm wrong.

>> oversee the hordes of AI agents

This is the thing I don't really get. I enjoy tinkering with AI and seeing what it comes up with to solve problems. But when I need to write working code that does anything beyond simple CRUD, it's faster for me to write the code than it is to (1) describe the problem in English with sufficient detail and working theory, then (2) check the AI's work, understand what it's written, de-duplicate and dry it out.

I guess if I skipped step 2, it might save time, but it would be completely irresponsible to put it into production, so that's not an option in any world where I maintain code quality and the trust of my clients.

Plus, having AI code mixed into my projects also leaves me with an uneasy sense of being less able to diagnose future bugs. Yes, I still know where everything is, but I don't know it as well as if I'd written it myself. So I find myself going back and re-reviewing AI-written code, re-familiarizing myself with it, in order to be sure I still have a full handle on everything.

To the extent that it may save me time as an engineer, I don't mind using it. But the degree to which the evangelists can peddle it to the management of a company as a replacement for human coders seems highly correlated with whether that company's management understood the value of safe code in the first place. If they didn't, then their infrastructure may have already been garbage, but it will now become increasingly unusable garbage. At some point, I think there will be a backlash when the results in reality can no longer be denied, and engineers who can come in and clean up the mess will be in high demand. But maybe that's just wishful thinking.

the top % of talent is still extremely hard to get, perhaps moreso

saw an article recently where every sector is seeing a reduction in IT/devs except for tech and ai companies

if your company is in a sector where eng is a cost-center and the product is not directly tied to your engineers / your company is pushing for efficiency it's an employer's market

OpenAI is implying that code may no longer be human readable in some circumstances.

> The resulting code does not always match human stylistic preferences, and that’s okay. As long as the output is correct, maintainable, and legible *to future agent runs*, it meets the bar.

https://openai.com/index/harness-engineering/

> But do the humans need to actually understand the code? A "yes" means the bottleneck is understanding (code review, code inspection). A "no" means you can go faster, but at some risk.

I always thought of things like code reviews as semi pseudo-science in most cases. I've sat through meetings where developers obviously understand the code that they are reviewing, but where they didn't understand anything about the system as a whole. If your perfect function pulls on 800 external dependencies that you trust. Trust because it's too much of a hazzle to go through them. I'd argue that in this situation you don't understand your code at all. I don't think it matters and I certainly don't think I'm better than anyone else in this regard. I only know how things work when it matters.

If anything, I think AI will increase human understanding without the need to write computer unfriendly code like "Clean Code", "DRY" and so on.

True, however as these products have been designed and coded by LLMs from the ground up in 2025+, they are generally using modern (typed even) languages, the latest version of third party libraries, usually have documentation of sorts... sometimes they even have test suites.

As such, they can often be improved as easily as one can prompt, which is much faster and easier than before. Notably in the FOSS world where one had to ask the maintainer, get ghosted for a year and have them go back with a "close: wontfix (too tedious)".

But the amount of pleasing useful art has gone up x1000; If I had a blog, I would now have access to art that would be a perfect fit for my words, whereas 5 years ago, I would have to do with a my own (talent-lacking) doodles.

Would some people prefer no art/illustration to AI generated art? Sure. But even more would prefer no art to my doodles.

its going to replace outsourcing to offshore teams pretty robustly though I suspect

Because it has been trained on Java class spaghetti.

When it’s trained on enough APL/K code, you’ll get minimal abstraction.

I’ve used Claude to write custom firmware in Rust for an ESP32-driven desktop clock.

Turned it into a Stripe revenue dashboard and notifier.

Even bought a couple more, flashed them, and gave to my cofounders, complete with AI written (personally tested, though) setup instructions!

It's already happening, particularly with "Ladybird Browser adopts Rust" [0] being at the top of HN today. It's now feasible to quickly iterate on a system's design with a dynamic language like Python, and then, once you're happy with the design, have AI rewrite it into something like Rust or Zig. I can even foresee a future where we intentionally maintain two parallel implementations, with machine-defined translation between them, such that we're able to do massive changes on the higher level implementation in minutes, and then once we finish iterating, have it run overnight to reimplement (or rewrite) it in the performant language. A bit like the difference between a unoptimized debugging version of a project, and the highly optimized one, but on steroids.

[0] https://news.ycombinator.com/item?id=47120899

Worth reiterating due to the skyrocketing costs of RAM.

The sad reality is it will likely be the older languages (I tend to see Ruby vibed a lot) just because there is so much more to train on.

With a bit of AI sprinkled in, Rust code can surely also waste gigabytes of RAM on "Hello World" ;)

Precisely, especially because habits have been bound up in the high (and difficult to measure!) cost of code. We got precious about it, really.

Not necessary: stand up meetings.

But that's the thing - changing course is suddenly no longer hard. We've already reached a state where I can have AI generate a decent set of tests from an existing codebase (or better yet, I'd already have them ahead of time), and to then do a massive refactoring or even a full rewrite while I get a good night's sleep. There is nothing "has always been" about this.

I believe the ChatGPT code has a bug, in that it accepts three spaces or tabs before a code fence, while the Google Markdown spec says up to three spaces, and does not allow a tab there.

I also see that the tests generated by ChatGPT are far too few for the code features implemented. The cannot be the result of actual red/green TDD where the test comes before the feature is added.

For examples, 1) the code allows "~~~" but only tests for "```", 2) there are no tests when len(fence) < fence_len nor when len(fence) > fence_len, and 3) there are no tests for leading spaces.

There's also duplicate code. The function _strip_closing_hashes is used once, in the line:

  text = _strip_closing_hashes(m.group("text")).strip()

The function is:

  def _strip_closing_hashes(s: str) -> str:
      s = s.rstrip()
      # remove trailing " ###" style closers
      s = re.sub(r"[ \t]+#+\s*$", "", s).rstrip()
      return s

The ".rstrip()" is unneeded as the ".strip()" does both lstrip and rstrip.

I think that rstrip() should be replaced with a strip(), the function renamed to "_get_inline_content", and used as "text = _get_inline_content(m.group("text")).

Also, the Google spec also says "A sequence of # characters with anything but spaces following it is not a closing sequence, but counts as part of the contents of the heading:" so is it really correct to use "\s*" in that regex, instead of "[ ]*"? And does it matter, since the input was rstrip'ped already?

So perhaps:

  def _get_inline_content(s: str) -> str:
      s = s.rstrip(" ") # remove trailing spaces
      s = s.rstrip("#") # removing "#" style closers
      return s.strip() # remove leading and trailing whitespace

would be more correct, readable, and maintainable?

Exactly. I feel like the latest models are basically a couple MCP servers away from just doing the whole thing. You just say, here’s what I want the system to do, and it’ll just do everything. No knowledge required. You need only know how to ask.

I wonder if they will make code fast in the way that McDonalds made food fast. For many business needs, knowing when a project will finish would be equally or even more valuable than knowing that it will contain more code or employ fewer programmers.

The Airbus A320neo can already takeoff, ascend, cruise, descend, and land all by autopilot. It can even download your flight plan from the airline's servers.

But you still need the pilots because the system can only handle the happy path. As soon as there's any blockade or strong weather change, the autopilot will just turn off. And then you need the pilots.

I would say software engineering with AI is similar: The AI can handle CRUD just fine. But once things get messy, you need someone who can actually think.

To fly a plane with 300+ passengers you still only need 2-3 pilots. That has remained consistent with the invention of autopilot. While we might still need a few human engineer experts, maybe we only need a few for small to medium sized companies? That may not eliminate the career for the top % but it effectively does for the vast majority of engineers.

We do automate lots about flying, not just take-off and landing. It's why a 4-engine aircraft in the 1960s required flight crews of 6-8 people just to fly the thing when they can be routinely flown with 2-3 today.

Autolands absolutely do exist.

See my heading half way down the page!

> the interesting shift is where the time goes. before: thinking + typing. now: thinking + reviewing.

It's widely accepted that you can't learn just by reading, you have to write. So only thinking and reviewing is a great way to lose all the business domain knowledge.

> the thinking part didn't get cheaper -- domain knowledge, edge cases, integration constraints -- none of that is free. what changed is you now review AI output instead of type your own, which is genuinely faster but not as different as it sounds

It's very different - you lose business domain knowledge if you're only reading.

Have you tried Gemini 3.1 Pro for that yet? They put a ton of work into improving its frontend abilities.

Yeah it's more important, but I think that has become a whole lot cheaper too.