7 ms·

Whenever I get worried about this I comb through our ticket tracker and see that ~0% of them can be implemented by AI as it exists today. Once somebody cracks t

Whenever I get worried about this I comb through our ticket tracker and see that ~0% of them can be implemented by AI as it exists today. Once somebody cracks the memory problem and ships an agent that progressively understands the business and the codebase, then I'll start worrying. But context limitation is fundamental to the technology in its current form and the value of SWEs is to turn the bigger picture into a functioning product.

malyk 8mo ago
Can you give an example to help us understand? I look at my ticket tracker and I see basically 100% of it that can be done by AI. Some with assistance because business logic is more complex/not well factored than it should be, but most of the work that is done AI is perfectly capable of doing with a well defined prompt.
- contagiousflow 8mo ago
  Why do you have a backlog then? If a current AI can do 100% of it then just run it over the weekend and close everything
  - fishpham 8mo ago
    As always, the limit is human bandwidth. But that's basically what AI-forward companies are doing now. I would be curious which tasks OP commenter has that couldn't be done by an agent (assuming they're a SWE)
    - Analemma_ 8mo ago
      This sounds bogus to me: if AI really could close 100% of your backlog with just a couple more humans in the loop, you’d hire a bunch of temps/contractors to do that, then declare the product done and lay off everybody. How come that isn’t happening?
      - fishpham 8mo ago
        Because there's an unlimited amount of work to do. This is the same reason you are not fired once completing a feature :-) The point of hiring a FTE is to continue to create work that provides business value. For your analogy, FTEs often do that by hiring temp, and you can think of the agent as the new temp in this case - the human drives an infinite amount of them
        sarchertech 8mo ago
        Why hasn’t any of the software I use started shipping features at a breakneck speed then? The only thing any of them have added is barely working AI features. Why aren’t there 10x the number of games on steam? Why aren’t people releasing new integrated programming language/OS/dev environments? Why does our backlog look exactly the same as when I left for posterity leave 4 months ago?
        fishpham 8mo ago
        Questions posed in bad faith can only be answered by the author.
        sarchertech 8mo ago
        Someone asked why the backlog doesn’t get finished. You answered that it does but the backlog just refills. So I asked where is the backlog evidence that the original backlog was completed. I’m still waiting for the evidence. I still haven’t seen externally verifiable evidence that AI is a net productivity boost for the ability to ship software. That doesn’t mean that it isn’t. It does mean that it isn’t big enough to be obvious. I’m very closet watching every external metric I can find. Nothing yet. Just saw the steam metrics for January. Fewer titles than January last year.
        catmanjan 8mo ago
        Sounds more like busy work rather than something that makes money
- dwa3592 8mo ago
  >>I look at my ticket tracker and I see basically 100% of it that can be done by AI. That's a sign that you have spurious problems under those tickets or you have a PM problem. Also, a job is a not a task- if your company has jobs which is a single task then those jobs would definitely be gone.
- lbrito 8mo ago
  Then why isn't it? Just offload it to the clankers and go enjoy a margarita at the beach or something.
  - Gud 8mo ago
    There are plenty of people who are enjoying margarita by the beach while you, the laborer, are working for them.
    - lbrito 8mo ago
      Preach. That's always been the case though, AI just makes it slightly worse.
- rockbruno 8mo ago
  I think the "well defined prompt" is precisely what the person you responded to is alluring to. They are saying they don't get worried because AI doesn't get the job done without someone behind it that knows exactly what to prompt.
- gordonhart 8mo ago
  Here's an example ticket that I'll probably work on next week: Live stream validation results as they come in The body doesn't give much other than the high-level motivation from the person who filed the ticket. In order to implement this, you need to have a lot of context, some of which can be discovered by grepping through the code base and some of which can't: - What is the validation system and how does it work today? - What sort of UX do we want? What are the specific deficiencies in the current UX that we're trying to fix? - What prior art exists on the backend and frontend, and how much of that can/should be reused? - Are there any scaling or load considerations that need to be accounted for? I'll probably implement this as 2-3 PRs in a chain touching different parts of the codebase. GPT via Codex will write 80% of the code, and I'll cover the last 20% of polish. Throughout the process I'll prompt it in the right direction when it runs up against questions it can't answer, and check its assumptions about the right way to push this out. I'll make sure that the tests cover what we need them to and that the resultant UX feels good. I'll own the responsibility for covering load considerations and be on the line if anything falls over. Does it look like software engineering from 3 years ago? Absolutely not. But it's software engineering all the same even if I'm not writing most of the code anymore.
  - Rodeoclash 8mo ago
    This right here is my view on the future as well. Will the AI write the entire feature in one go? No. Will the AI be involved in writing a large proportion of the code that will be carefully studied and adjusted by a human before being used? Absolutely yes. This cyborg process is exactly how we're using AI in our organisation as well. The human in the loop understands the full context of what the feature is and what we're trying to achieve.
  - fragmede 8mo ago
    I mean, what is the validation system? Either it exists in code, and thus can be discovered if you point the AI at repo, or... what, it doesn't exist? For the UX, have it explore your existing repos and copy prior art from there and industry standards to come up with something workable. Web scale issues can be inferred by the rest of the codebase. If your terraform repo has one RDS server, vs a fleet of them, multi-region, then the AI, just as well as a human, can figure out if it needs Google Spanner level engineering or not. (probably not) Bigger picture though, what's the process of a human logs an under specified ticket and someone else picks it up and has no clue what to do with it? They're gonna go ask the person who logged the bug for their thoughts and some details beyond "hurr Durr something something validation". If we're at the point where AI is able to make a public blog post shaming the open source developer for not accepting a patch, throwing questions back to you in JIRA about the details of the streaming validation system is well within its capabilities, given the right set of tools.
ninetyninenine 8mo ago
[dead]
matt_heimer 8mo ago
It's not binary. Jobs will be lost because management will expect the fewer developers to accomplish more by leveraging AI.
- louiereederson 8mo ago
  Big tech might ahead of the rest of the economy in this experiment. Microsoft grew headcount by ~3% from June 2022 to June 2025 while revenue grew by >40%. This is admittedly weak anecdata but my subjective experience is their products seem to be crumbling (GitHub problems around the Azure migration for instance), and worse than they even were before. We'll see how they handle hiring over the next few years and if that reveals anything.
  - JetSpiegel 8mo ago
    Well, Google just raised prices by 30% on the GSuite "due to AI value delivered", but you can't even opt out, so even revenue is a bullshit metric.
- datsci_est_2015 8mo ago
  Already built in. We haven’t hired recently and our developers are engaged in a Cold War to set the new standard of productivity.
e_i_pi_2 8mo ago
A lot of this can be provided or built up by better documentation in the codebase, or functional requirements that can also be created, reviewed, and then used for additional context. In our current codebase it's definitely an issue to get an AI "onboarded", but I've seen a lot less hand-holding needed in projects where you have the AI building from the beginning and leaving notes for itself to read later
- tharkun__ 8mo ago
  We have this in some of our projects too but I always wonder how long it's going to take until it just fails. Nobody reads all those memory files for accuracy. And knowing what kind of BS the AI spews regularly in day to day use I bet this simply doesn't scale.
- gordonhart 8mo ago
  Curious to hear if you've seen this work with 100k+ LoC codebases (i.e. what you could expect at a job). I've had some good experiences with high autonomy agents in smaller codebases and simpler systems but the coherency starts to fizzle out when the system gets complicated enough that thinking it through is the hard part as opposed to hammering out the code.
  - christkv 8mo ago
    Our codebase is well over 250k and we have a hierarchy of notes for the modules so we read as much as we need for the job with a base memory that explains how the notes work
  - wenc 8mo ago
    If you vector index your code base, agents can explore it without loading it into context. This is what Cursor and Roo and Kiro and probably others do. Claude Code uses string searches. What helps is also getting it to generate a docs of your code so that it has map. This is actually how humans understand a large code base too. We don’t hold a large code base in memory — we navigate it through docs and sampling bits of code.
  - 8mo ago
rockbruno 8mo ago
While true, my personal fear is that the higher-ups will overlook this fact and just assume that AI can do everything because of some cherry-pick simple examples, leading to one of those situations where a bunch of people get fired for no reason and then re-hired again after some time.
- palmotea 8mo ago
  > leading to one of those situations where a bunch of people get fired for no reason and then re-hired again after some time. More likely they get fired for no reason, never rehired, and the people left get burned out trying to hold it all together.
  - easymodex 8mo ago
    Exactly, now which one do you wanna be? The burned out ones but still working in SWE or the fired ones which in the long run converge to manual labor which AI can't do. Not to mention in SWE case the salaries would be pushed down to match cost of AI doing it.
- themafia 8mo ago
  As if "higher-ups" is an assigned position. If you fail as a "higher up" you're no longer higher up. Then someone else can take your place. To the extent this does not naturally happen is evidence of petty or major corruptions within the system.
  - Seattle3503 8mo ago
    In competitive industries, bad firms will fail. Some industries are not competitive though. I have a friend that went a little crazy working as a PM at a large health insurance firm.
danesparza 8mo ago
Apparently you haven't seen ChatGPT enterprise and codex. I have bad news for you ...
- gordonhart 8mo ago
  Codex with their flagship model (currently GPT-5.3-Codex) is my daily driver. I still end up doing a lot of steering!
pupppet 8mo ago
We're all slowly but surely lowering our standards as AI bombards us with low-quality slop. AI doesn't need to get better, we all just need to keep collectively lowering our expectations until they finally meet what AI can currently do, and then pink-slips away.
- tines 8mo ago
  Exactly. This happens in every aspect of life. Something convenient comes along and people will accommodate it despite it being worse, because people don’t care.
zozbot234 8mo ago
> Once somebody cracks the memory problem and ships an agent that progressively understands the business and the codebase, then I'll start worrying. Um, you do realize that "the memory" is just a text file (or a bunch of interlinked text files) written in plain English. You can write these things out yourself. This is how you use AI effectively, by playing to its strengths and not expecting it to have a crystal ball.
deet 8mo ago
Just keep in mind that there are many highly motivated people directly working on this problem. It's hard to predict how quickly it will be solved and by whom first, but this appears to be a software engineering problem solvable through effort and resources and time, not a fundamental physical law that must be circumvented like a physical sciences problem. Betting it won't be solved enough to have an impact on the work of today relatively quickly is betting against substantial resources and investment.
- datsci_est_2015 8mo ago
  The implication of your assertion is pretty much a digital singularity. You’re implying that there will be no need for humans to interact with the digital world at all, because any work in the digital world will be achievable by AI. Wonder what that means for meatspace. Edit: Would also disagree this isn’t a physics problem. Pretty sure power required scales according to problem complexity. At a certain level of problem complexity we’re pretty much required to put enough carbon in the atmosphere to cook everyone to a crisp. Edit 2: illustrative example, an Epic in Jira: “Design fusion reactor”
- slopinthebag 8mo ago
  Why do you think it's not a physical sciences problem? It could be the case that current technologies simply cannot scale due to fundamental physical issues. It could even be a fundamental rule of intelligent life, that one cannot create intelligence that surpasses its own. Plenty of things get substantial resources and investment and go nowhere. Of course I could be totally wrong and it's solved in the next couple years, it's almost impossible to make these predictions either way. But I get the feeling people are underestimating what it takes to be truly intelligent, especially when efficiency is important.
  - jatari 8mo ago
    >It could even be a fundamental rule of intelligent life, that one cannot create intelligence that surpasses its own. Well that is easily disproved by the fact that people have children with higher IQ's than their own.
nemo1618 8mo ago
"The steamroller is still many inches away. I'll make a plan once it actually starts crushing my toes." You are in danger. Unless you estimate the odds of a breakthrough at <5%, or you already have enough money to retire, or you expect that AI will usher in enough prosperity that your job will be irrelevant, it is straight-up irresponsible to forgo making a contingency plan.
- Gigachad 8mo ago
  My contingency plan is that if AI leaves me unable to get a job, we are all fucked and society as a whole will have to fix the situation and if it doesn’t, there is nothing I could have done about it anyway.
  - everettde 8mo ago
    this has been me ever since my philosophy undergrad.
  - tired-turtle 8mo ago
    This is a sensible plan, given your username.
    - nikkwong 8mo ago
      Yeah seriously. Don't people understand the fact that society is not good at mopping up messes like this—there has been a K shaped economy for several decades now and most Americans have something like $400 in their bank accounts. The bottom had already fallen out for them, and help still hasn't arrived. I think it's more likely that what really happens is that white collar workers, especially the ones on the margin, join this pool—and there is a lot of suffering for a long time. Personally, rather devolving into nihilism, I'd rather try to hedge against suffering that fate. Now is the time to invest and save money. (or yesterday)
      - sarchertech 8mo ago
        If white collar workers as a whole suffer severe economic setback over a short term timespan, your savings and investments won’t help you. Unless you’re investing in guns, ammo, food, and a bunker. We’re talking worse unemployment than depression era Germany. And structurally more significant unemployment because the people losing their jobs were formally very high earners.
sensanaty 8mo ago
I look through the backlog for my team consisting of 9 trillion ill-defined (if defined at all) tickets that tells you basically nothing. The large, overwhelming majority of my team's time is spent on combing through these tickets and making sense of them. Once we know what the ticket is even trying to say, we're usually out with the solution in a few days at most, so implementation isn't the bottleneck, nowhere near. This scenario has been the same everywhere I've ever worked, at large, old institutions as well as fresh startups. The day I'll start worrying is when the AI is capable of following the web of people involved to translate what the vaguely phrased ticket that's been backlogged for God knows how long actually means
- arcologies1985 8mo ago
  At my workplace we now use Claude Code to parse written specs and source code, search through JIRA, and draft, refine and organize tickets (using the JIRA API via a CLI tool). Way faster than through the UI. However as you point out we have no program-accessible source of data on who stakeholders, contributors, managers, etc. are and have to write a lot of that ourselves. For a smaller business perhaps one could write all of that down in an accessible way to improve this but for a large dynamic business it seems very difficult.
yodsanklai 8mo ago
> ~0% of them can be implemented by AI as it exists today I think it's more nuanced than that. I'd say that - 0% can't be implemented by AI - but a lot of them can be implemented much faster thanks to AI - a lot of them can be implemented slower when using AI (because author has to fix hallucinations, revert changes that caused bugs) As we learn to use these tools, even in their current state, they will increase productivity by some factor and reduce needs for programmers.
- sarchertech 8mo ago
  What factor of increased productivity will lead to reduced need for programmers? I have seen numerous 25-50% productivity boosts over my career. Not a single one of them reduced the overall need for programmers. I can’t even think of one that reduced the absolute number of programmers in a specific field.
UncleOxidant 8mo ago
The memory problem is already being addressed in various ways - antigravity seems to keep a series of status/progress files describing what's been done, what needs doing, etc. A bit clunky, but it seems to work - I can open it up on a repo that I was working in a few days back and it seems to pick up this context such that I don't have to completely bring it up to speed every time like I used to have to do. I've heard that claude code has similar mechanisms. I've been doing stuff with recent models (gemini 3, claude 4.5/6, even smaller, open models like GLM5 and Qwen3-coder-next) that was just unthinkable a few months back. Compiler stuff, including implementing optimizations, generating code to target a new, custom processor, etc. I can ask for a significant new optimization feature in our compiler before going to lunch and come back to find it implemented and tested. This is a compiler that targets a custom processor so there is also verilog code involved. We're having the AI make improvements on both the hardware and software sides - this is deep-in-the-weeds complex stuff and AI is starting to handle it with ease. There are getting to be fewer and fewer things in the ticket tracker that AI can't implement. A few months ago I would've completely agreed with you, but the game is changing very rapidly now.
- taysco 8mo ago
  this works fine for like 2-3 small instruction sets. once you start getting to scale of a real enterprise system, the AI falls down and can't handle that amount of context. It will start ignoring critical pieces or not remember them. And without constant review AI will start priotizing things that are not your business priority. I don't agree they have solved this problem, at all, or really in any way that's actually usable.
  - UncleOxidant 8mo ago
    What I'm saying is, don't get to thinking that the memory problem is some kind of insurmountable, permanent barrier that's going to keep us safe. It's already being addressed, maybe crudely at first, but the situation is already much better than it was - I no longer have to bring the model up to speed completely every time I start a new session. Part of this is much larger context windows (1M tokens now). New architectures are also being proposed to deal with the issue, as well.
krackers 8mo ago
>progressively understands the business This is no different than onboarding a new member of the team, and I think openAI was working on that "frontier" >We started by looking at how enterprises already scale people. They create onboarding processes. They teach institutional knowledge and internal language. They allow learning through experience and improve performance through feedback. They grant access to the right systems and set boundaries. AI coworkers need the same things. And tribal knowledge will not be a moat once execs realize that all they need to do is prioritize documentation instead of "code velocity" as a metric (sure any metric gets goodhearted, but LLMs are great at sifting through garbage to find the high perplexity tokens). >But context limitation is fundamental to the technology in its current form This may not be the case, large enough context-windows plus external scratchpads would mostly obviate the need for true in context learning. The main issue today is that "agent harnesses" suck. The fact that claude code is considered good is more an indication of how bad everything else is. Tool traces read like a drunken newb brute-forcing his way through tasks. LLMs can mostly "one-shot" individual functions, but orchestrating everything is the blocker. (Yes there's progress in metr or whatever but I don't trust any of that, else we'd actually see the results in real-world open source projects). LLMs don't really know how to interact with subagents. They're generally sort of myopic even with tool calls. They'll spend 20 minutes trying to fix build issues going down a rabbit hole without stepping back to think. I think some sort of self-play might end up solving all of these things, they need to develop a "theory of mind" in the same way that humans do, to understand how to delegate and interact with the subagents they spawn. (Today a failure case is agents often don't realize subagents don't share the same context.) Some of this is certainly in the base model and pretraining, but it needs to be brought out in the same way RL was needed for tool use.
vincent_s 8mo ago
Ha, this triggered me. I'm building exactly this. It's a coding agent that takes a ticket from your tracker, does the work asynchronously, and replies with a pull request. It does progressively understand the codebase. There's a pre-warming step so it's already useful on the first ticket, but it gets better with each one it completes. The agent itself is done and working well. Right now I'm building out the infrastructure to offer it as a SaaS. If anyone wants to try it, hit me up. Email is in my profile. Website isn't live yet, but I'm putting together a waitlist.
audience_mem 8mo ago
0%? This is as wrong as people who say it can do 100% of tasks.

malyk 8mo ago

Can you give an example to help us understand? I look at my ticket tracker and I see basically 100% of it that can be done by AI. Some with assistance because business logic is more complex/not well factored than it should be, but most of the work that is done AI is perfectly capable of doing with a well defined prompt.

contagiousflow 8mo ago
Why do you have a backlog then? If a current AI can do 100% of it then just run it over the weekend and close everything
- fishpham 8mo ago
  As always, the limit is human bandwidth. But that's basically what AI-forward companies are doing now. I would be curious which tasks OP commenter has that couldn't be done by an agent (assuming they're a SWE)
  - Analemma_ 8mo ago
    This sounds bogus to me: if AI really could close 100% of your backlog with just a couple more humans in the loop, you’d hire a bunch of temps/contractors to do that, then declare the product done and lay off everybody. How come that isn’t happening?
    - fishpham 8mo ago
      Because there's an unlimited amount of work to do. This is the same reason you are not fired once completing a feature :-) The point of hiring a FTE is to continue to create work that provides business value. For your analogy, FTEs often do that by hiring temp, and you can think of the agent as the new temp in this case - the human drives an infinite amount of them
      - sarchertech 8mo ago
        Why hasn’t any of the software I use started shipping features at a breakneck speed then? The only thing any of them have added is barely working AI features. Why aren’t there 10x the number of games on steam? Why aren’t people releasing new integrated programming language/OS/dev environments? Why does our backlog look exactly the same as when I left for posterity leave 4 months ago?
        fishpham 8mo ago
        Questions posed in bad faith can only be answered by the author.
        sarchertech 8mo ago
        Someone asked why the backlog doesn’t get finished. You answered that it does but the backlog just refills. So I asked where is the backlog evidence that the original backlog was completed. I’m still waiting for the evidence. I still haven’t seen externally verifiable evidence that AI is a net productivity boost for the ability to ship software. That doesn’t mean that it isn’t. It does mean that it isn’t big enough to be obvious. I’m very closet watching every external metric I can find. Nothing yet. Just saw the steam metrics for January. Fewer titles than January last year.
      - catmanjan 8mo ago
        Sounds more like busy work rather than something that makes money
dwa3592 8mo ago
>>I look at my ticket tracker and I see basically 100% of it that can be done by AI. That's a sign that you have spurious problems under those tickets or you have a PM problem. Also, a job is a not a task- if your company has jobs which is a single task then those jobs would definitely be gone.
lbrito 8mo ago
Then why isn't it? Just offload it to the clankers and go enjoy a margarita at the beach or something.
- Gud 8mo ago
  There are plenty of people who are enjoying margarita by the beach while you, the laborer, are working for them.
  - lbrito 8mo ago
    Preach. That's always been the case though, AI just makes it slightly worse.
rockbruno 8mo ago
I think the "well defined prompt" is precisely what the person you responded to is alluring to. They are saying they don't get worried because AI doesn't get the job done without someone behind it that knows exactly what to prompt.
gordonhart 8mo ago
Here's an example ticket that I'll probably work on next week: Live stream validation results as they come in The body doesn't give much other than the high-level motivation from the person who filed the ticket. In order to implement this, you need to have a lot of context, some of which can be discovered by grepping through the code base and some of which can't: - What is the validation system and how does it work today? - What sort of UX do we want? What are the specific deficiencies in the current UX that we're trying to fix? - What prior art exists on the backend and frontend, and how much of that can/should be reused? - Are there any scaling or load considerations that need to be accounted for? I'll probably implement this as 2-3 PRs in a chain touching different parts of the codebase. GPT via Codex will write 80% of the code, and I'll cover the last 20% of polish. Throughout the process I'll prompt it in the right direction when it runs up against questions it can't answer, and check its assumptions about the right way to push this out. I'll make sure that the tests cover what we need them to and that the resultant UX feels good. I'll own the responsibility for covering load considerations and be on the line if anything falls over. Does it look like software engineering from 3 years ago? Absolutely not. But it's software engineering all the same even if I'm not writing most of the code anymore.
- Rodeoclash 8mo ago
  This right here is my view on the future as well. Will the AI write the entire feature in one go? No. Will the AI be involved in writing a large proportion of the code that will be carefully studied and adjusted by a human before being used? Absolutely yes. This cyborg process is exactly how we're using AI in our organisation as well. The human in the loop understands the full context of what the feature is and what we're trying to achieve.
- fragmede 8mo ago
  I mean, what is the validation system? Either it exists in code, and thus can be discovered if you point the AI at repo, or... what, it doesn't exist? For the UX, have it explore your existing repos and copy prior art from there and industry standards to come up with something workable. Web scale issues can be inferred by the rest of the codebase. If your terraform repo has one RDS server, vs a fleet of them, multi-region, then the AI, just as well as a human, can figure out if it needs Google Spanner level engineering or not. (probably not) Bigger picture though, what's the process of a human logs an under specified ticket and someone else picks it up and has no clue what to do with it? They're gonna go ask the person who logged the bug for their thoughts and some details beyond "hurr Durr something something validation". If we're at the point where AI is able to make a public blog post shaming the open source developer for not accepting a patch, throwing questions back to you in JIRA about the details of the streaming validation system is well within its capabilities, given the right set of tools.

ninetyninenine 8mo ago

[dead]

matt_heimer 8mo ago

It's not binary. Jobs will be lost because management will expect the fewer developers to accomplish more by leveraging AI.

louiereederson 8mo ago
Big tech might ahead of the rest of the economy in this experiment. Microsoft grew headcount by ~3% from June 2022 to June 2025 while revenue grew by >40%. This is admittedly weak anecdata but my subjective experience is their products seem to be crumbling (GitHub problems around the Azure migration for instance), and worse than they even were before. We'll see how they handle hiring over the next few years and if that reveals anything.
- JetSpiegel 8mo ago
  Well, Google just raised prices by 30% on the GSuite "due to AI value delivered", but you can't even opt out, so even revenue is a bullshit metric.
datsci_est_2015 8mo ago
Already built in. We haven’t hired recently and our developers are engaged in a Cold War to set the new standard of productivity.

e_i_pi_2 8mo ago

A lot of this can be provided or built up by better documentation in the codebase, or functional requirements that can also be created, reviewed, and then used for additional context. In our current codebase it's definitely an issue to get an AI "onboarded", but I've seen a lot less hand-holding needed in projects where you have the AI building from the beginning and leaving notes for itself to read later

tharkun__ 8mo ago
We have this in some of our projects too but I always wonder how long it's going to take until it just fails. Nobody reads all those memory files for accuracy. And knowing what kind of BS the AI spews regularly in day to day use I bet this simply doesn't scale.
gordonhart 8mo ago
Curious to hear if you've seen this work with 100k+ LoC codebases (i.e. what you could expect at a job). I've had some good experiences with high autonomy agents in smaller codebases and simpler systems but the coherency starts to fizzle out when the system gets complicated enough that thinking it through is the hard part as opposed to hammering out the code.
- christkv 8mo ago
  Our codebase is well over 250k and we have a hierarchy of notes for the modules so we read as much as we need for the job with a base memory that explains how the notes work
- wenc 8mo ago
  If you vector index your code base, agents can explore it without loading it into context. This is what Cursor and Roo and Kiro and probably others do. Claude Code uses string searches. What helps is also getting it to generate a docs of your code so that it has map. This is actually how humans understand a large code base too. We don’t hold a large code base in memory — we navigate it through docs and sampling bits of code.
- 8mo ago

rockbruno 8mo ago

While true, my personal fear is that the higher-ups will overlook this fact and just assume that AI can do everything because of some cherry-pick simple examples, leading to one of those situations where a bunch of people get fired for no reason and then re-hired again after some time.

palmotea 8mo ago
> leading to one of those situations where a bunch of people get fired for no reason and then re-hired again after some time. More likely they get fired for no reason, never rehired, and the people left get burned out trying to hold it all together.
- easymodex 8mo ago
  Exactly, now which one do you wanna be? The burned out ones but still working in SWE or the fired ones which in the long run converge to manual labor which AI can't do. Not to mention in SWE case the salaries would be pushed down to match cost of AI doing it.
themafia 8mo ago
As if "higher-ups" is an assigned position. If you fail as a "higher up" you're no longer higher up. Then someone else can take your place. To the extent this does not naturally happen is evidence of petty or major corruptions within the system.
- Seattle3503 8mo ago
  In competitive industries, bad firms will fail. Some industries are not competitive though. I have a friend that went a little crazy working as a PM at a large health insurance firm.

danesparza 8mo ago

Apparently you haven't seen ChatGPT enterprise and codex. I have bad news for you ...

gordonhart 8mo ago
Codex with their flagship model (currently GPT-5.3-Codex) is my daily driver. I still end up doing a lot of steering!

pupppet 8mo ago

We're all slowly but surely lowering our standards as AI bombards us with low-quality slop. AI doesn't need to get better, we all just need to keep collectively lowering our expectations until they finally meet what AI can currently do, and then pink-slips away.

tines 8mo ago
Exactly. This happens in every aspect of life. Something convenient comes along and people will accommodate it despite it being worse, because people don’t care.

zozbot234 8mo ago

> Once somebody cracks the memory problem and ships an agent that progressively understands the business and the codebase, then I'll start worrying. Um, you do realize that "the memory" is just a text file (or a bunch of interlinked text files) written in plain English. You can write these things out yourself. This is how you use AI effectively, by playing to its strengths and not expecting it to have a crystal ball.

deet 8mo ago

Just keep in mind that there are many highly motivated people directly working on this problem. It's hard to predict how quickly it will be solved and by whom first, but this appears to be a software engineering problem solvable through effort and resources and time, not a fundamental physical law that must be circumvented like a physical sciences problem. Betting it won't be solved enough to have an impact on the work of today relatively quickly is betting against substantial resources and investment.

datsci_est_2015 8mo ago
The implication of your assertion is pretty much a digital singularity. You’re implying that there will be no need for humans to interact with the digital world at all, because any work in the digital world will be achievable by AI. Wonder what that means for meatspace. Edit: Would also disagree this isn’t a physics problem. Pretty sure power required scales according to problem complexity. At a certain level of problem complexity we’re pretty much required to put enough carbon in the atmosphere to cook everyone to a crisp. Edit 2: illustrative example, an Epic in Jira: “Design fusion reactor”
slopinthebag 8mo ago
Why do you think it's not a physical sciences problem? It could be the case that current technologies simply cannot scale due to fundamental physical issues. It could even be a fundamental rule of intelligent life, that one cannot create intelligence that surpasses its own. Plenty of things get substantial resources and investment and go nowhere. Of course I could be totally wrong and it's solved in the next couple years, it's almost impossible to make these predictions either way. But I get the feeling people are underestimating what it takes to be truly intelligent, especially when efficiency is important.
- jatari 8mo ago
  >It could even be a fundamental rule of intelligent life, that one cannot create intelligence that surpasses its own. Well that is easily disproved by the fact that people have children with higher IQ's than their own.

nemo1618 8mo ago

"The steamroller is still many inches away. I'll make a plan once it actually starts crushing my toes." You are in danger. Unless you estimate the odds of a breakthrough at <5%, or you already have enough money to retire, or you expect that AI will usher in enough prosperity that your job will be irrelevant, it is straight-up irresponsible to forgo making a contingency plan.

Gigachad 8mo ago
My contingency plan is that if AI leaves me unable to get a job, we are all fucked and society as a whole will have to fix the situation and if it doesn’t, there is nothing I could have done about it anyway.
- everettde 8mo ago
  this has been me ever since my philosophy undergrad.
- tired-turtle 8mo ago
  This is a sensible plan, given your username.
  - nikkwong 8mo ago
    Yeah seriously. Don't people understand the fact that society is not good at mopping up messes like this—there has been a K shaped economy for several decades now and most Americans have something like $400 in their bank accounts. The bottom had already fallen out for them, and help still hasn't arrived. I think it's more likely that what really happens is that white collar workers, especially the ones on the margin, join this pool—and there is a lot of suffering for a long time. Personally, rather devolving into nihilism, I'd rather try to hedge against suffering that fate. Now is the time to invest and save money. (or yesterday)
    - sarchertech 8mo ago
      If white collar workers as a whole suffer severe economic setback over a short term timespan, your savings and investments won’t help you. Unless you’re investing in guns, ammo, food, and a bunker. We’re talking worse unemployment than depression era Germany. And structurally more significant unemployment because the people losing their jobs were formally very high earners.

sensanaty 8mo ago

I look through the backlog for my team consisting of 9 trillion ill-defined (if defined at all) tickets that tells you basically nothing. The large, overwhelming majority of my team's time is spent on combing through these tickets and making sense of them. Once we know what the ticket is even trying to say, we're usually out with the solution in a few days at most, so implementation isn't the bottleneck, nowhere near. This scenario has been the same everywhere I've ever worked, at large, old institutions as well as fresh startups. The day I'll start worrying is when the AI is capable of following the web of people involved to translate what the vaguely phrased ticket that's been backlogged for God knows how long actually means

arcologies1985 8mo ago
At my workplace we now use Claude Code to parse written specs and source code, search through JIRA, and draft, refine and organize tickets (using the JIRA API via a CLI tool). Way faster than through the UI. However as you point out we have no program-accessible source of data on who stakeholders, contributors, managers, etc. are and have to write a lot of that ourselves. For a smaller business perhaps one could write all of that down in an accessible way to improve this but for a large dynamic business it seems very difficult.

yodsanklai 8mo ago

> ~0% of them can be implemented by AI as it exists today I think it's more nuanced than that. I'd say that - 0% can't be implemented by AI - but a lot of them can be implemented much faster thanks to AI - a lot of them can be implemented slower when using AI (because author has to fix hallucinations, revert changes that caused bugs) As we learn to use these tools, even in their current state, they will increase productivity by some factor and reduce needs for programmers.

sarchertech 8mo ago
What factor of increased productivity will lead to reduced need for programmers? I have seen numerous 25-50% productivity boosts over my career. Not a single one of them reduced the overall need for programmers. I can’t even think of one that reduced the absolute number of programmers in a specific field.

UncleOxidant 8mo ago

The memory problem is already being addressed in various ways - antigravity seems to keep a series of status/progress files describing what's been done, what needs doing, etc. A bit clunky, but it seems to work - I can open it up on a repo that I was working in a few days back and it seems to pick up this context such that I don't have to completely bring it up to speed every time like I used to have to do. I've heard that claude code has similar mechanisms. I've been doing stuff with recent models (gemini 3, claude 4.5/6, even smaller, open models like GLM5 and Qwen3-coder-next) that was just unthinkable a few months back. Compiler stuff, including implementing optimizations, generating code to target a new, custom processor, etc. I can ask for a significant new optimization feature in our compiler before going to lunch and come back to find it implemented and tested. This is a compiler that targets a custom processor so there is also verilog code involved. We're having the AI make improvements on both the hardware and software sides - this is deep-in-the-weeds complex stuff and AI is starting to handle it with ease. There are getting to be fewer and fewer things in the ticket tracker that AI can't implement. A few months ago I would've completely agreed with you, but the game is changing very rapidly now.

taysco 8mo ago
this works fine for like 2-3 small instruction sets. once you start getting to scale of a real enterprise system, the AI falls down and can't handle that amount of context. It will start ignoring critical pieces or not remember them. And without constant review AI will start priotizing things that are not your business priority. I don't agree they have solved this problem, at all, or really in any way that's actually usable.
- UncleOxidant 8mo ago
  What I'm saying is, don't get to thinking that the memory problem is some kind of insurmountable, permanent barrier that's going to keep us safe. It's already being addressed, maybe crudely at first, but the situation is already much better than it was - I no longer have to bring the model up to speed completely every time I start a new session. Part of this is much larger context windows (1M tokens now). New architectures are also being proposed to deal with the issue, as well.

krackers 8mo ago

>progressively understands the business This is no different than onboarding a new member of the team, and I think openAI was working on that "frontier" >We started by looking at how enterprises already scale people. They create onboarding processes. They teach institutional knowledge and internal language. They allow learning through experience and improve performance through feedback. They grant access to the right systems and set boundaries. AI coworkers need the same things. And tribal knowledge will not be a moat once execs realize that all they need to do is prioritize documentation instead of "code velocity" as a metric (sure any metric gets goodhearted, but LLMs are great at sifting through garbage to find the high perplexity tokens). >But context limitation is fundamental to the technology in its current form This may not be the case, large enough context-windows plus external scratchpads would mostly obviate the need for true in context learning. The main issue today is that "agent harnesses" suck. The fact that claude code is considered good is more an indication of how bad everything else is. Tool traces read like a drunken newb brute-forcing his way through tasks. LLMs can mostly "one-shot" individual functions, but orchestrating everything is the blocker. (Yes there's progress in metr or whatever but I don't trust any of that, else we'd actually see the results in real-world open source projects). LLMs don't really know how to interact with subagents. They're generally sort of myopic even with tool calls. They'll spend 20 minutes trying to fix build issues going down a rabbit hole without stepping back to think. I think some sort of self-play might end up solving all of these things, they need to develop a "theory of mind" in the same way that humans do, to understand how to delegate and interact with the subagents they spawn. (Today a failure case is agents often don't realize subagents don't share the same context.) Some of this is certainly in the base model and pretraining, but it needs to be brought out in the same way RL was needed for tool use.

vincent_s 8mo ago

Ha, this triggered me. I'm building exactly this. It's a coding agent that takes a ticket from your tracker, does the work asynchronously, and replies with a pull request. It does progressively understand the codebase. There's a pre-warming step so it's already useful on the first ticket, but it gets better with each one it completes. The agent itself is done and working well. Right now I'm building out the infrastructure to offer it as a SaaS. If anyone wants to try it, hit me up. Email is in my profile. Website isn't live yet, but I'm putting together a waitlist.

audience_mem 8mo ago

0%? This is as wrong as people who say it can do 100% of tasks.