The consequences of task switching in supervisory programming
90 points
1 day ago
| 7 comments
| martinfowler.com
| HN
simonw
7 hours ago
[-]
This cognitive debt bit from the linked article by Margaret-Anne Storey at https://margaretstorey.com/blog/2026/02/09/cognitive-debt/ is fantastic:

> But by weeks 7 or 8, one team hit a wall. They could no longer make even simple changes without breaking something unexpected. When I met with them, the team initially blamed technical debt: messy code, poor architecture, hurried implementations. But as we dug deeper, the real problem emerged: no one on the team could explain why certain design decisions had been made or how different parts of the system were supposed to work together. The code might have been messy, but the bigger issue was that the theory of the system, their shared understanding, had fragmented or disappeared entirely. They had accumulated cognitive debt faster than technical debt, and it paralyzed them.

reply
appplication
7 hours ago
[-]
This was essentially my experience vibe coding a web app. I got great results initially and made it quite far quickly but over time velocity exponentially slowed due to exactly this cognitive debt. Took my time and did a ground up rewrite manually and made way faster progress and a much more stable app.

You could argue LLMs let me learn enough about the product I was trying to build that the second rewrite was faster and better informed, and that’s probably true to some degree, but it also was quite a few weeks down the drain.

reply
Mavvie
5 hours ago
[-]
That makes sense, but surely there's a middle ground somewhere between "AI does everything including architecture" and writing everything by hand?
reply
appplication
5 hours ago
[-]
Of course! The original attempt wasn’t really AI doing everything. I was writing much of the code but letting AI drive general patterns since I was unfamiliar with web dev. Now, it’s also not entirely without AI, but I am very much steering the ship and my usage of AI is more “low context chat” than “agentic”. IMO it’s a more functional way to interface with AI for anyone with solid engineering skills.
reply
r_lee
5 hours ago
[-]
I think the sweet spot is to make the initial stuff yourself and then extend or modify somewhat with LLMs

it acts as a guide for the LLM too, so it doesn't have to just come up with everything on its own in terms of style or design choices in terms of consistency I'd say?

reply
elcritch
3 hours ago
[-]
For more complex projects I find this pattern very helpful. The last two gens of SOTA models have become rather good at following existing code patterns.

If you have a solid architecture they can be almost prescient in their ability to modify things. However they're a bit like Taylor series expansions. They only accurate out so far from the known basis. Hmm, or control theory where you have stable and unstable regimes.

reply
mattmanser
2 hours ago
[-]
I think it's closer to "doing everything by hand" than you'd expect.

For me, anyway.

I design as I code, the architecture becomes more obvious as I fill in the detail.

So getting AI to do bits, really means getting AI to do the really easy bits.

reply
svara
2 hours ago
[-]
> So getting AI to do bits, really means getting AI to do the really easy bits.

As someone who gets quickly bored with repetitive work, this is big though.

reply
ReptileMan
2 hours ago
[-]
I spend half my prompts of making codex explain why and what is he doing. The other 40% is reducing the size of the code base and optimizing. Only 10-sh percent is new development.
reply
phren0logy
7 hours ago
[-]
I'm not a coder, I'm a medical doctor. I see some interesting parallels in how medical students sort themselves into specialties by cognitive style to this new rift in programming with LLMs.

Some people like the deep work, some like managing a steady rain of chaos. There's no one right answer. But I'll tell you that my classmates who are happy as nephrologists are very different to the ones that are happy as transplant surgeons.

reply
wittlesus
7 hours ago
[-]
The cognitive debt framing really resonates. I've noticed two very different failure modes when working with AI coding tools:

1. The "black box" problem — you accept generated code without fully understanding it, and later can't debug or extend it because you never built the mental model.

2. The "context fragmentation" problem — you're constantly switching between reviewing AI output, correcting course, and holding your own design intent in your head. It's like pair programming with someone who has perfect syntax recall but zero memory of what you discussed 5 minutes ago.

The teams that seem to avoid the wall are the ones who treat AI as a drafting tool rather than a decision-making tool. You still need to be the architect. The AI just types faster than you do.

reply
evmar
9 hours ago
[-]
The term he’s searching for is possibly “intellectual control”: https://www.georgefairbanks.com/ieee-software-v36-n1-jan-201...
reply
themafia
8 hours ago
[-]
> a third of them were instantly converted to being very pro-LLM. That suggests that practical experience

I wasn't aware one could get 'practical experience' "instantly." I would assume that their instant change of heart owes more to other factors. Perhaps concern over the source of their next paycheck? You have admitted you just "forced" them to do this. Isn't the question then, why didn't they do it before? Shouldn't you answer that before you prognosticate?

> that junior developers will still be needed, if nothing else because they are open-minded about LLMs

You're broadcasting, to me, that you understand all of the above perfectly, yet instead of acknowledging it, you're planning on taking advantage of it.

> I think the equivalent of cruft is ignorance

Exceedingly ironic.

> Will two-pizza teams shrink to one-pizza teams

The language you use to describe work, workers, and overcoming challenges are too depressing to continue. You have become everything we hated about this profession.

reply
simonw
7 hours ago
[-]
If you haven't experenced a post-November-2025 coding agent before and someone coaches you through how to one-shot prompt it into solving a difficult problem in your own codebase that you are deeply familiar with I can see how you might be an almost instant convert.

(Based on your comment history I'm guessing you haven't experienced this yourself yet.)

reply
svara
2 hours ago
[-]
You're right, and I enjoy using coding agents too. I've built some things with them I wouldn't have otherwise.

However, it's been a full quarter now since November 2025.

Based on facts on the ground, i.e. the rate and quality of new software and features we observe, change has been nowhere as dramatic as your comment would suggest.

It seems to me that a possible explanation is that people get very excited about massive speedups in specific tasks, but the bottleneck of the system shifts somewhere else immediately (e.g, human capacity for learning, team coordination costs, communication delays).

reply
icedchai
6 hours ago
[-]
This rings true for me. Up until the end of 2025 I had my doubts. I haven't fully adopted AI, but I am using it for several side projects where I normally would not have made much progress. The output w/Claude Code is solid.
reply
themafia
4 hours ago
[-]
The challenges I have were selected because I enjoy solving them and because very few, if any, people have taken the time to work on them already. As such I have no desire to "one-shot" a solution and I additionally have serious doubts that any model trained on any existing code could possibly output anything useful or anything that truly fits into the design of the system. These projects are written for style and are to explore ideas and gain experience. Inviting an LLM in out of laziness is completely the opposite of my intentions.

The only other code that I write is for a handful of industry specific products that are not challenging in any way to code but are fun to design for the specific needs of my users and are informed by their incredible feedback from the field. The time and effort to play games with an LLM prompt would have effectively zero value here and again is the opposite of what makes these products great enough to be sold by word of mouth alone.

Aside from all of this I have no desire to pay a subscription to a service that requires me to submit all of my code to their engine for output. Given their models apparent fondness for taking copyrighted code and passing it off as it's own I would not put it past them to play games behind my back with my work.

Finally I see no new "AI billionaires" suddenly rising out of the field and I see no "AI heavy" companies suddenly increasing their profits, productivity or quality in any way. I hear what you are saying, and you're certainly not alone in saying it, but I see zero evidence that it's actually meaningful in the real world software market.

reply
wiz21c
2 hours ago
[-]
I would be very happy to solve problems that "very few, if any, people have taken the time to work on them already."

My experience (as someone how works with a team of PhD's) is that code is about 30% of what we do but in these, 75% are "trivial things" (building charts, quickly designing apps to process information, etc). Out of these 75%, AI certainly helps us at least 50% of the time (and amazes me 10% of the time :-))

> I see no new "AI billionaires" suddenly rising out of the field and I see no "AI heavy" companies suddenly increasing their profits, productivity or quality in any way.

Exactly what I was telling myself yesterday. That's rather not in line with the media coverage.

reply
habinero
7 hours ago
[-]
I have literally heard this exact vague phrase about every single stupid model that has come out, plus more than a few companies.

So far it's all been endless unfounded FOMO hype by people who have something to sell or podcasts to be on. I am so tired of it.

reply
simonw
7 hours ago
[-]
Ask around and see if you can find anyone you know who's experienced the November 2025 effect. Claude Code / Codex with GPT-5.1+ or Opus 4.5+ really did make a material difference - they flipped the script from "can write code that often works" to "can write code that almost always works".

I know you'll dismiss that as the same old crap you've heard before, but it's pretty widely observed now.

reply
geraneum
4 hours ago
[-]
I’ve been living this experience and using latest models in work throughout this time. The failure modes of LLMs have not fundamentally changed. The makers are not awfully transparent about what exactly they change in each model release the same way you know what changed in i.e., a new Django version. But there’s not been a paradigm shift. I believe/guess (from outside) the big change you think you’re experiencing could be result of many things like better post training processes (RLHF) for models to run a predefined set of commands like always running tests, or other marginal improvements to the models and focusing on programming tasks. To be clear these improvements are welcome and useful, just not the groundbreaking change some claim.
reply
ej88
4 hours ago
[-]
the perimeter of the tasks the LLMs can handle continuously expands at a pretty steady pace

a year ago they could easily one shot full stack features in my hobby next.js apps but imploded in my work codebase

as of opus 4.6 they can now one shot full features in a complex js/go data streaming & analysis tool but implode in low latency voice synthesis systems (...for now...)

just depends on how you're using it (skill issues are a thing) and what you're working on

reply
kalessin
10 hours ago
[-]
I like the idea of "cognitive debt" vs "technical debt".
reply
jongjong
11 hours ago
[-]
Part of me feels like LLMs will struggle to architect code properly, no matter how good they get.

Software engineering is different from programming. Other kinds of engineers often ridiculed software engineers as "not real engineers" because mainstream engineers never had to build arbitrarily complex software systems from scratch. They have never experienced the cascading issues which often happen when trying to make changes to complex software systems. Their brief exposure to programming during their university days gave them a glimpse into programming but not software engineering. They think they understand it but they don't.

Other engineers think that they're the only ones wrestling with the laws of nature.

They're wrong. Software engineering involves wrestling with entropy itself. In some ways, it's an even purer form of engineering. Software engineering struggles against the most fundamental forces and requires reasoning skills of the highest order.

I think software engineers will be among the last of the white collar professions to be automated besides the ones which have legal protections like lawyers, judges, politicians, accountants, pilots... where a human is required to provide a layer of accountability. Though I think lawyers will be reduced to being "official human stamping machines" before software engineers are reduced to mere Product Owners.

reply
Swizec
10 hours ago
[-]
> Though I think lawyers will be reduced to being "official human stamping machines" before software engineers are reduced to mere Product Owners

GeLLMan Amnesia – AI can fully automate every profession except the ones I’m deeply familiar with.

I’m a software engineer who wears the product owner hat a lot these days, there’s no way AI will automate this any time soon. Too much peopling and accountability.

reply
Herring
9 hours ago
[-]
Don't be so sure about that. These days I'm already finding it 100x easier/informative to have complicated charged discussions (eg immigration) with Gemini than with actual people. It's day and night. Accountability might be solvable too, maybe escrow and pay me if you waste my time. Or amazon-like reviews.
reply
Aurornis
8 hours ago
[-]
> These days I'm already finding it 100x easier/informative to have complicated charged discussions (eg immigration) with Gemini than with actual people.

It’s scary how quickly people start to mistake LLMs appeasing them for actual conversation.

Discussing something with an LLM isn’t equivalent to having a conversation with a person. It’s just a text generator trained to show you what you want to see.

reply
Herring
8 hours ago
[-]
Don't assume everyone is like you. I'm an early adopter and I know how to scaffold it. I generally love the bleeding edge in all things, and I'm increasingly sure it's an actual talent to be able to quickly adapt to unfamiliar things (this includes not making assumptions).
reply
paulryanrogers
8 hours ago
[-]
What do you feel like you get out of discussions with Gemini about politically charged topics like immigration?
reply
Herring
8 hours ago
[-]
I think that's obvious from the discussion so far? It broadens my horizons.

YMMV. Ask the bot for supporting evidence, and follow up on google/wikipedia.

reply
KittenInABox
8 hours ago
[-]
Would you be willing to share the logs of a nuanced conversation you've had with Gemini?
reply
Herring
7 hours ago
[-]
I'm not sure. Most of it is not even on the logs, it's followed up elsewhere.

You can try something like this on Gemini 3 Pro:

> Break down aspects of the economy by amenability to state control high/medium/low, based on what we see in successful economies. Include a rationale and supporting evidence/counterexamples. Present it in 3 tables.

It should give you dozens of things you can look up. It might mention successful Singapore and Vienna-style public housing. Some nice videos on that on Youtube.

Online discussions are usually at the level of "[Flagged] Communism bad".

reply
linkregister
2 hours ago
[-]
I have the luxury of a few friends capable of discussing complex military, political, and social issues who are able to hold nuanced views backed by evidence.

Because of that good fortune, it hasn't occurred to me to use an LLM to organize information for these topics. I appreciate your sharing your approach and I look forward to trying this use case of LLMs.

reply
metadat
10 hours ago
[-]
With the requisite planning steps Codex and Claude are already coming up with better architecture and design than I can.

I've been doing this for more than 25 years.

reply
whattheheckheck
9 hours ago
[-]
What's the most complicated thing its designed so far for you
reply
Herring
9 hours ago
[-]
Not that guy, but for me it's something like Tensorflow/Pytorch. A domain-specific language for a scientific application, Python API with a Rust core for very fast/safe calculations. It has all kinds of bells & whistles you'd want, like automatic differentiation, lazy evaluation, provenance, serialization, etc. Occasionally dips down to raw pointer work too. It's easy to test, so AI excels at this type of thing.
reply
sebmellen
11 hours ago
[-]
Beautifully expressed… you missed doctors in your list of white collar professions, but I’m sure surgeons and pilots will outlive all of us from an AI resilience standpoint.
reply
jongjong
10 hours ago
[-]
Ah yes 100%, doctors have legal moats too.

It's kind of terrifying to think that all professions are going to have to shift away from value creation to pure politics to survive.

I have a feeling that big tech companies will be legally forced to pay royalties to software engineers. Once software engineers stop applying their reasoning skills to solving real problems and start vengefully focusing it on politics, we're going to corrupt the whole system in our favor. We have enough collective knowledge to frame such corruption as moral in the context of an already corrupt system.

Either software engineers will create regulatory moats for themselves or there will be a more broad political movement like communism. I've met many people working deep in the critical systems which underpin our society who are full-blown communists.

reply
whattheheckheck
9 hours ago
[-]
Openclaw as your own person political advisor is pretty cool
reply