This is a real thing. I spent all of January doing Greenfield development using Claude (I finished the requirements) and all I can say is thank goodness I had the Max 5x plan and not the 20x as I got breaks once the tokens were used up till the next cycle. I was forced to get up and do something else. That something else was biking, rowing, walking. My productivity had never been higher but at what cost? My health no thanks. So I'm glad I'm using the time till token reset for my health. I time it perfectly. I do a walk, row, bike for 1 hour then as I arrive back the tokens are reset. I get like 3 hours nonstop use per token batch with the 5x plan. I've been thinking about going 20x but am scared...
Hypothesis: limiting usage / tokens could have a positive effect on project quality, since it forces the developer to think more carefully about the problems they're working on. When you're forced to stop and slow down, you try to be more deliberate with token usage. But if you have unlimited tokens you can just keep generating infinite lines of code without thinking as hard about the problem.
I've seen people on social media bragging about how they're able to produce a mountain of code as if this was praiseworthy.
I don’t get this tbh, I use Claude too and my issue is the opposite - too many small breaks. Every time I hit enter my brain wants to checkout because the agent just spins while it creates thousands of tokens and churns on the subject. Even if it’s only 2m, that’s 2m where my mind has nothing to work on.
Hard to stay in flow and engaged.
Feels weirdly similar to being interrupted over slack.
you are correct flow is not achieved as this is not programming more like system design, architecture, QA, Product Owner work. It's using the swarm as your own dev team.
> It's using the swarm as your own dev team.
reply
Managing high performance dev/ops teams is it's own form of a state of flow. In fact for me, it's much more addicting than any other as the outcomes are usually many multiples of any IC role you could have. Even crazier when you have a "follow the sun" team involved so there the work just gets sequentially handed off and is always in constant motion.
I imagine AI coding is like this for a lot of folks.
But it's also programming as you have to study outputs to ensure they're correct. Some (it seems many) don't do this, and then their outputs usually aren't correct.
In my experience, QA is something like ensuring it responds correctly to input. This is similar, but not the same as code review. I would more liken QA to dynamic review rather than static. Note though that code review can still be a form of QA. (Formal proofing especially.)
That’s what QA departments in software companies do. In many other contexts they examine things produced by machines to ensure they meet the specs and functional requirements for that piece, and if not, either adjust it, have someone else adjust it, or have the adjusted machine spit out another one. They might design tests, fixtures to measure things, etc etc etc but they do not make the things directly.
To be fair, ensuring that machines produce the correct outputs (even by making someone else fix it) is still the kind of process I'm talking about. After all, that's also how it works in software.
yes agreed. I'm running 3-5 parallel Claude at once with requirements as the input. My prompt is say work on section 5.1 or something very specific. Then I'm monitoring the work across all instances.
Sorry, I wasn’t giving a serious answer. It’s just not as amusingly worded as in thought.
Seriously, though, your question is one of those “how long is a piece of string” sort of questions. Just like any other software quality question, it depends on context, competence, goals, market dynamics, organizational culture, project timelines, team expertise, etc.
Do people pay their bills on time? Do people wear seatbelts? Do people brush their teeth for the full two recommended minutes? Depends, depends, depends.
Sorry, I didn't realise you weren't the OP. I was really asking the OP as they said they had large productivity gains from using AI to code. But if you're a professional developer, the same question can be answered by you: do you specifically review all AI generated production code?
In my own case 100% of my code is reviewed by humans (generally me), and that IMO is the only sensible option, and has been the standard since I started coding commercially 33 years ago. I don't use AI to generate code though, other than a few experiments, as I don't really need to write much code these days.
When people talk about AI increasing developer productivity, they usually focus on the coding part.
In my experience, the bigger change happens after the code is written.
When you move from writing code to supervising agents, your output increases — but your cognitive load increases too.
Instead of writing every line yourself, you're now monitoring systems:
Did the agent go off-script?
Did it retry 50 times while I was asleep?
What did that run actually cost?
The strange part is that the mental burden doesn't disappear just because the agent is autonomous.
In some ways it gets worse, because failures become harder to notice early and harder to contain once they start.
It starts to feel less like programming and more like running operations for a team of extremely fast, extremely literal junior developers.
Curious if others are seeing the same shift.
That really sounds like micro managing jr. developers.
I wonder if the interface for this kind of thing might be better presented as a sort of JIRA ticket system. Define a dependency graph of work with the ability to break down any ticket into more tickets or change priority or relationships etc.
Though I think the micro manage part still doesn’t fit into that model. You’d need the code-level view and not just a ticket covering the tests that satisfy the spec and
performance goals.
I think a lot of people feel this tension.
Programming used to be mostly about building things directly.
You write code, run it, fix it, repeat.
With agents it starts to shift toward supervision:
define the task, watch the output, correct the drift.
It's a different kind of work.
Sometimes it feels less like programming and more like
managing a very fast team that never gets tired but
also never really understands the goal unless you
spell it out extremely carefully.
I suspect a lot of developers still enjoy the "building"
part more than the "supervising" part.
> That really sounds like micro managing jr. developers.
That's how I tend to describe AI to a lot of non-technical people (I actually generally say it's like having an really fresh intern who can read technical docs insanely fast but needs a lot of supervision).
Yeah I am not sure many people gonna hang around this - I am not sure I wanna do this role. I like building and delivering and ai is great help but I will not be happy supervising agents, there are better jobs. Unless the money is not to be refused
They're probably talking about some point after the capabilities of LLMs started to become clear.
It's why Codex, Claude Code, Gemini CLI etc. were developed at all - it was clear that if you wanted a concrete application of LLMs with clear productivity benefits, coding was low-hanging fruit, so all the AI vendors jumped on that and started hyping it.
Sure, but jumping from its amazing these things work for code at all to software engineering is solved is something only grifters or those drunk on the kool-aid did.
I do agree that it was thought that these llm-agents would be extremely useful and that is why they were developed, and I happen to believe they in fact are extremely useful (without disagreeing that much of the stuff in the article definitely does happen.)
I just sort of resent the setup that it was supposed to be X but actually it failed, when not only is there only minor evidence that it failed, but it was only a brief period in time when it was supposed to be X.
Prior to the rise of LLM coding, developers had to, from time to time, spent time deep diving through large amounts of code they didn't write. Hundreds of thousands of lines or even millions. This might happens when starting a new job, or changing projects, or when tasked with evaluating some third party tech or integrating it, or taking over something that was previously owned by another developer.
During such phases of work, it's not unusual to put in some long hours in order to get up to speed.
With LLMs, it is possible to perpetually experience hundreds of thousands of lines of third party code, on about a weekly basis.
But this is not the same. The code is not known by anyone, anywhere else. It exists nowhere else, and so has no track record of deployment. No documentation, nothing. It's not something where you can concentrate on making a small modification, while trusting the rest of it to be working.
I use it every day and I'm taking off weekends for the first time in a decade. It's done wonders for my mental health. I think teams should pay more attention to the value of pumping the brakes vs. incessant redlining. We may actually be able to have a healthy relationship with AI then.
Selection bias? The early adopters that are motivated to adopt tools to deliver more, typically also were working more to start with and may have already been struggling with their rate of output?
> I'm 8x more productive than I was in 2022... And I jokingly say "I'm probably not going to have a job in 1 or 2 years"...
> We are going to create incredible value to humanity. 8x rate. I don't know what our hourly will be.
Do we have 8x more demand for software than current developers can currently supply? No, we don’t. Many developers will soon have a very bitter pill to swallow once they realize that developers are the beneficiaries of good market dynamics rather than some precious intellectual elite whose skills are monetizable in any other context.
Their hourly will be whatever DoorDash pays them to deliver pizza and egg rolls. Grad school isn’t any safer, and frankly, many of the soft, arrogant and maladroit people I’ve seen try to enter blue collar trades fail very quickly once they realize how hard the road is to get to those high salaries they always hear about.
I feel totally the opposite. I feel like I'm better able to have more work-life balance. Our predictions are more accurate. I'm enjoying working on actual problems rather than boilerplate. These tools are amazing
Personally, I make a lot more "out of hour" commits than I used to because I'll batch up low priority tasks throughout the day and let the computer chug on them at night when I'm elsewhere. Commits are coming in at all hours, but I'm not actually looking at them until the next morning.
1. llms allow devs to be more productive, so more free time is seen as opportunity for more work. ppl overshoot and just work more
2. generalized tooling makes devs seem more replaceable putting downward pressure on job security (ie work harder or we’ll get someone who will, oh and for less money)
3. llms allow for more “multitasking” (debatable) via many running background tasks, so more opportunities to “just finish one more thing”
In my experience, code validation (unit testing, code review, manual testing, etc.) was more of a bottleneck than producing code for the most part. This means that faster code generation wouldn't produce significant gains in throughput unless the code validation speeds up too. In my workplace, I've seen evidence that the people showing the biggest productivity gains from AI coding are now shipping enormous commits that are barely getting any validation. Given the Zeitgeist, others are for some reason more lenient towards that than they normally would be (or should be).
I can't deny that this might be a trend in practice, but at companies with reasonably self-aware practices, it isn't, or doesn't need to be.
There's this weird thing that happens with new tools where people seem to surrender their autonomy to them, e.g. "welp, I just get pings from [Slack|my phone|etc] all the time, nothing I can do than just be interrupted constantly." More recently, it's "this failed because Claude chose..." No, Claude didn't choose, the person who submitted the PR chose to accept it.
It's possible to use tools responsibly and effectively. It's also possible to encourage and mentor employees to do that. The idea that a dev has to be effectively on call because they're pushing AI slop is just wrong on so many levels.
> It's possible to use tools responsibly and effectively. It's also possible to encourage and mentor employees to do that.
It's not in the company's interest to stop employees from overworking. Having people overwork for the same pay under pressure is the desired outcome, actually.
There are many companies which don’t operate like this. What you’re describing is the rather uniquely American ultracapitalist perspective. But as an employee, you have a choice to work for a company like that, or work for one that’s run by humans.
Well if the management want to get more AI, they gonna get more AI, and no, I am not gonna be running around making sure their dreams work smoothly under my human supervision - I am gonna let it go all the way they want. In the mean time I focus on improving my skills.
A developers job has always been reviewing and understanding code.
Code is literally always the last resort. Unless you're building solutions for other customer, most companies should attempt to minimise the amount of code they have. Because, and I repeat, it's a developers job to understand and review code. More code, more understanding needed, more reviews needed, more problems created.
This is a real thing. I spent all of January doing Greenfield development using Claude (I finished the requirements) and all I can say is thank goodness I had the Max 5x plan and not the 20x as I got breaks once the tokens were used up till the next cycle. I was forced to get up and do something else. That something else was biking, rowing, walking. My productivity had never been higher but at what cost? My health no thanks. So I'm glad I'm using the time till token reset for my health. I time it perfectly. I do a walk, row, bike for 1 hour then as I arrive back the tokens are reset. I get like 3 hours nonstop use per token batch with the 5x plan. I've been thinking about going 20x but am scared...
Hypothesis: limiting usage / tokens could have a positive effect on project quality, since it forces the developer to think more carefully about the problems they're working on. When you're forced to stop and slow down, you try to be more deliberate with token usage. But if you have unlimited tokens you can just keep generating infinite lines of code without thinking as hard about the problem.
I've seen people on social media bragging about how they're able to produce a mountain of code as if this was praiseworthy.
One might wonder if the trend holds when limiting token use to… zero?
I don’t get this tbh, I use Claude too and my issue is the opposite - too many small breaks. Every time I hit enter my brain wants to checkout because the agent just spins while it creates thousands of tokens and churns on the subject. Even if it’s only 2m, that’s 2m where my mind has nothing to work on.
Hard to stay in flow and engaged.
Feels weirdly similar to being interrupted over slack.
you are correct flow is not achieved as this is not programming more like system design, architecture, QA, Product Owner work. It's using the swarm as your own dev team.
> It's using the swarm as your own dev team. reply
Managing high performance dev/ops teams is it's own form of a state of flow. In fact for me, it's much more addicting than any other as the outcomes are usually many multiples of any IC role you could have. Even crazier when you have a "follow the sun" team involved so there the work just gets sequentially handed off and is always in constant motion.
I imagine AI coding is like this for a lot of folks.
But it's also programming as you have to study outputs to ensure they're correct. Some (it seems many) don't do this, and then their outputs usually aren't correct.
That’s what my teammates are for, I pipe slack and jira to Claude and the asker and teammates tell me if there’s a bug
Sounds more like code-level QA to me.
In my experience, QA is something like ensuring it responds correctly to input. This is similar, but not the same as code review. I would more liken QA to dynamic review rather than static. Note though that code review can still be a form of QA. (Formal proofing especially.)
That’s what QA departments in software companies do. In many other contexts they examine things produced by machines to ensure they meet the specs and functional requirements for that piece, and if not, either adjust it, have someone else adjust it, or have the adjusted machine spit out another one. They might design tests, fixtures to measure things, etc etc etc but they do not make the things directly.
To be fair, ensuring that machines produce the correct outputs (even by making someone else fix it) is still the kind of process I'm talking about. After all, that's also how it works in software.
I have never been in a flow state with an agent running. I use agents, but that isn’t flow.
and flow state is a luxury in 2026 with AI swarm most likely to be found sparingly if all. Good luck all!
yes agreed. I'm running 3-5 parallel Claude at once with requirements as the input. My prompt is say work on section 5.1 or something very specific. Then I'm monitoring the work across all instances.
I have similar problem but I have to switch contexts and it makes the work a lot more intense.
Are you a single agent user?
At least in my case, flow is gone. It’s all context switching now.
Sounds like a recipe for burnout
This. And another problem is that I feel not proud after completing the task. No sense of achievement.
Does a person review all the AI generated code?
Not at all unless they’re a) competently b) making something worth anything at all that c) isn’t a proof-of-concept or the like.
Yes, of course. I mean, is all production code reviewed?
Sorry, I wasn’t giving a serious answer. It’s just not as amusingly worded as in thought.
Seriously, though, your question is one of those “how long is a piece of string” sort of questions. Just like any other software quality question, it depends on context, competence, goals, market dynamics, organizational culture, project timelines, team expertise, etc.
Do people pay their bills on time? Do people wear seatbelts? Do people brush their teeth for the full two recommended minutes? Depends, depends, depends.
Sorry, I didn't realise you weren't the OP. I was really asking the OP as they said they had large productivity gains from using AI to code. But if you're a professional developer, the same question can be answered by you: do you specifically review all AI generated production code?
In my own case 100% of my code is reviewed by humans (generally me), and that IMO is the only sensible option, and has been the standard since I started coding commercially 33 years ago. I don't use AI to generate code though, other than a few experiments, as I don't really need to write much code these days.
Great shilling attempt )
When people talk about AI increasing developer productivity, they usually focus on the coding part. In my experience, the bigger change happens after the code is written. When you move from writing code to supervising agents, your output increases — but your cognitive load increases too. Instead of writing every line yourself, you're now monitoring systems: Did the agent go off-script? Did it retry 50 times while I was asleep? What did that run actually cost? The strange part is that the mental burden doesn't disappear just because the agent is autonomous. In some ways it gets worse, because failures become harder to notice early and harder to contain once they start. It starts to feel less like programming and more like running operations for a team of extremely fast, extremely literal junior developers. Curious if others are seeing the same shift.
That really sounds like micro managing jr. developers.
I wonder if the interface for this kind of thing might be better presented as a sort of JIRA ticket system. Define a dependency graph of work with the ability to break down any ticket into more tickets or change priority or relationships etc.
Though I think the micro manage part still doesn’t fit into that model. You’d need the code-level view and not just a ticket covering the tests that satisfy the spec and performance goals.
I think a lot of people feel this tension. Programming used to be mostly about building things directly. You write code, run it, fix it, repeat. With agents it starts to shift toward supervision: define the task, watch the output, correct the drift. It's a different kind of work. Sometimes it feels less like programming and more like managing a very fast team that never gets tired but also never really understands the goal unless you spell it out extremely carefully. I suspect a lot of developers still enjoy the "building" part more than the "supervising" part.
> That really sounds like micro managing jr. developers.
That's how I tend to describe AI to a lot of non-technical people (I actually generally say it's like having an really fresh intern who can read technical docs insanely fast but needs a lot of supervision).
Yeah I am not sure many people gonna hang around this - I am not sure I wanna do this role. I like building and delivering and ai is great help but I will not be happy supervising agents, there are better jobs. Unless the money is not to be refused
> Software engineering was supposed to be artificial intelligence’s easiest win.
At what point in time? Did anyone foresee coding being one of the best and soonest applications of this stuff?
I seem to recall short snippets of IDE code completion being one of the first commercial applications of it.
No one saw it coming.
They're probably talking about some point after the capabilities of LLMs started to become clear.
It's why Codex, Claude Code, Gemini CLI etc. were developed at all - it was clear that if you wanted a concrete application of LLMs with clear productivity benefits, coding was low-hanging fruit, so all the AI vendors jumped on that and started hyping it.
Sure, but jumping from its amazing these things work for code at all to software engineering is solved is something only grifters or those drunk on the kool-aid did.
I do agree that it was thought that these llm-agents would be extremely useful and that is why they were developed, and I happen to believe they in fact are extremely useful (without disagreeing that much of the stuff in the article definitely does happen.)
I just sort of resent the setup that it was supposed to be X but actually it failed, when not only is there only minor evidence that it failed, but it was only a brief period in time when it was supposed to be X.
Because swe was the furthest advanced "collaborative cognition" field in terms of human workflows
[dead]
Prior to the rise of LLM coding, developers had to, from time to time, spent time deep diving through large amounts of code they didn't write. Hundreds of thousands of lines or even millions. This might happens when starting a new job, or changing projects, or when tasked with evaluating some third party tech or integrating it, or taking over something that was previously owned by another developer.
During such phases of work, it's not unusual to put in some long hours in order to get up to speed.
With LLMs, it is possible to perpetually experience hundreds of thousands of lines of third party code, on about a weekly basis.
But this is not the same. The code is not known by anyone, anywhere else. It exists nowhere else, and so has no track record of deployment. No documentation, nothing. It's not something where you can concentrate on making a small modification, while trusting the rest of it to be working.
I use it every day and I'm taking off weekends for the first time in a decade. It's done wonders for my mental health. I think teams should pay more attention to the value of pumping the brakes vs. incessant redlining. We may actually be able to have a healthy relationship with AI then.
Selection bias? The early adopters that are motivated to adopt tools to deliver more, typically also were working more to start with and may have already been struggling with their rate of output?
We've become cashiers.
My 6 year old is doing my job.
The best I can hope for is that HN article that said the word "Context".
I know the magic words "Make me a single page html js web app"... or "Install Virtual Box with Fedora Cinnamon using CLI"....
I'm 8x more productive than I was in 2022... And I jokingly say "I'm probably not going to have a job in 1 or 2 years"...
We are going to create incredible value to humanity. 8x rate. I don't know what our hourly will be.
> I'm 8x more productive than I was in 2022... And I jokingly say "I'm probably not going to have a job in 1 or 2 years"...
> We are going to create incredible value to humanity. 8x rate. I don't know what our hourly will be.
Do we have 8x more demand for software than current developers can currently supply? No, we don’t. Many developers will soon have a very bitter pill to swallow once they realize that developers are the beneficiaries of good market dynamics rather than some precious intellectual elite whose skills are monetizable in any other context.
Their hourly will be whatever DoorDash pays them to deliver pizza and egg rolls. Grad school isn’t any safer, and frankly, many of the soft, arrogant and maladroit people I’ve seen try to enter blue collar trades fail very quickly once they realize how hard the road is to get to those high salaries they always hear about.
I feel totally the opposite. I feel like I'm better able to have more work-life balance. Our predictions are more accurate. I'm enjoying working on actual problems rather than boilerplate. These tools are amazing
Personally, I make a lot more "out of hour" commits than I used to because I'll batch up low priority tasks throughout the day and let the computer chug on them at night when I'm elsewhere. Commits are coming in at all hours, but I'm not actually looking at them until the next morning.
two unthought out thoughts:
1. llms allow devs to be more productive, so more free time is seen as opportunity for more work. ppl overshoot and just work more
2. generalized tooling makes devs seem more replaceable putting downward pressure on job security (ie work harder or we’ll get someone who will, oh and for less money)
3. llms allow for more “multitasking” (debatable) via many running background tasks, so more opportunities to “just finish one more thing”
No silver bullet. We've known this since at least the 1980s. The fact that the authors of the code might not be human doesn't change this.
thouroughly reviewing and especially testing is faster than skipping manual review and tests
I'm just curious how much of this AI generated code is reviewed by humans at all, and if that is factored into the productivity gains.
In my experience, code validation (unit testing, code review, manual testing, etc.) was more of a bottleneck than producing code for the most part. This means that faster code generation wouldn't produce significant gains in throughput unless the code validation speeds up too. In my workplace, I've seen evidence that the people showing the biggest productivity gains from AI coding are now shipping enormous commits that are barely getting any validation. Given the Zeitgeist, others are for some reason more lenient towards that than they normally would be (or should be).
I can't deny that this might be a trend in practice, but at companies with reasonably self-aware practices, it isn't, or doesn't need to be.
There's this weird thing that happens with new tools where people seem to surrender their autonomy to them, e.g. "welp, I just get pings from [Slack|my phone|etc] all the time, nothing I can do than just be interrupted constantly." More recently, it's "this failed because Claude chose..." No, Claude didn't choose, the person who submitted the PR chose to accept it.
It's possible to use tools responsibly and effectively. It's also possible to encourage and mentor employees to do that. The idea that a dev has to be effectively on call because they're pushing AI slop is just wrong on so many levels.
> It's possible to use tools responsibly and effectively. It's also possible to encourage and mentor employees to do that.
It's not in the company's interest to stop employees from overworking. Having people overwork for the same pay under pressure is the desired outcome, actually.
There are many companies which don’t operate like this. What you’re describing is the rather uniquely American ultracapitalist perspective. But as an employee, you have a choice to work for a company like that, or work for one that’s run by humans.
> More recently, it's "this failed because Claude chose..." No, Claude didn't choose, the person who submitted the PR chose to accept it.
I can relate to this, unfortunately these tools are becoming a very convenient way to offload any kind of responsibility when something goes wrong.
Well if the management want to get more AI, they gonna get more AI, and no, I am not gonna be running around making sure their dreams work smoothly under my human supervision - I am gonna let it go all the way they want. In the mean time I focus on improving my skills.
[dead]
Was this comment written by an LLM? It seems like it was to me. e.g. the “paradox” is not a paradox at all and is just an obvious statement.
From looking at their other comments: unquestionably yes it's a bot.
I just assume any green account names are LLMs nowadays
A developers job has always been reviewing and understanding code.
Code is literally always the last resort. Unless you're building solutions for other customer, most companies should attempt to minimise the amount of code they have. Because, and I repeat, it's a developers job to understand and review code. More code, more understanding needed, more reviews needed, more problems created.
I don't know about you, but if I started doing all that instead of writing code as a priority, I'd be fired.
My job is to generate more money, not indulge in code.
Nah summarizing code is now an LLM job as well. There is no place for engineers in the new tech world order.
[dead]
[dead]