Post

That's the trade-off between intelligence and wisdom, I guess

I feel this unease, but I also feel that we’ll learn build cognitive scaffolding as we go. At least that’s what I’m telling myself about my next project

So eloquently put Simon, as always. I worked in Photography/Advertising for 15 years and stepped away because of whom I was becoming by working in that industry. It taught me to pay attention to the person the work creates, as much as the work the person creates. Context switching radiply=🧠🗑️.

these bots aren't good enough to be let off-leash. you get a big speedup, but not as much as it feels like. and utility is proportional to how much you pay for it

I think the right way to produce, verify, and keep a handle on generated code will be more and more of A Thing over time. That, more than anything, has me viscerally repulsed by "agent teams," Gas Town, other efforts to reduce the human input and review per LoC even more

This looks super interesting, thanks for sharing. I've been writing up super short PRDs in a docs folder of the repo, loosely connected into the original PRD.md. I've found that useful for keeping track of the "why".

I don’t know why I’m surprised to see this. I just thought was me — but it’s a very really problem that I’ve been struggling with.

tests

I've been trying a strategy against cognitive debt during a brownfield migration. A little mental model file. Always human edit. Agent checks for falseness only. I started calling it `HITL.md` and wrote it up. warrenwhipple.com/hitl

Writing about our experiences here: paragraph.com/@adamsobotka . Started as authoritative manifesto, ended as a messages from battlefield 😀

It's the classic "someone else wrote it and now I have to maintain it" problem sped up.

Yeah I have experienced the same. I think we should really spend more time to review and understand AI generated code. I find it helpful to write comments when reviewing the changes.

The debt isn't just in the code — it's in your ability to predict what breaks. You can review every line and still lose the intuition for why certain decisions were made. The fix is probably in the tooling: AI that annotates intent, not just implementation.

I love the "velocity without understanding is not sustainable" line from Mary-Anne's post. Pithy, accurate, and actionable: don't go faster than you can understand.

I've noticed this too. I've noticed that this is actually keeping me somewhat from using agents to their full capability, out of fear that I'd lose my mental model of what I'm (they're) building.

Super useful a have been collecting #cognitivedebt examples for an update on the below, from all kinds of fields but coding particularly good to pay attention to - like speedrunning the social effects we’ll maybe see elsewhere smithery.com/methods/cogn...

So the solution is not through humand being superior, but in humans being less collaborative. And that's why you prompt another AI conversation to review the claims combatively. There is no solidarity amongst the agents.

Another good one is "what tradeoffs did you make when writing this?" I have also prompted AI to write documentation for what it has built (pointing it at great examples of documentation elsewhere) and that has really helped for me.

Once we teach them to write documentation better, it will indeed make these systems more observable. We need to make them talk about the code in a novel way. Write lil O'Reily books about every vibe coded project.

I like to describe this as the race car analogy. AI-based coding is like driving a car at 150MPH, but to get round the track at a decent time, you have to slow down for the corners otherwise you crash. You can let the AI drive when the feature set is a straight road, but then you slow for corners

Institutional memory is the fix. CLAUDE.md + hooks + session protocol means the model carries context across sessions instead of you carrying it in your head. We wrote up how: https://www.dugganusa.com/post/the-online-consensus-is-wrong

yep. thanks for writing about this. definitely noticed similar patterns and wrote a bit about my own coping mechanisms galsapir.github.io/sparse-thoug...

I could definitely describe the high level architecture accurately, since that's where I've had to direct the LLM the most. But I am often surprised at the lower level details. Despite having read it all. I guess the gulf between doing the thing and reading about the thing is applicable to code too

Wrote about something similar in martin.janiczek.cz/2026/02/05/t... How do you keep familiarity with the code? A balance of writing code and letting agents write the code? Reviews don't quite do it for me, I seem to need to touch the code.

I wrote about this recently too, I think it is the biggest thing we will need to grapple with in agentic coding. I used a slightly different term (speed vertigo) but it's the same thing: joshtuddenham.dev/blog/vertigo/

“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” Pre-ai this happened with all software as OGs age out. The “why” of things is lost. Now takes months rather than years.

This reminds me of research debt distill.pub/2017/researc...

Thanks @grimalkina.bsky.social! Cognitive debt is unsurprising given the foundations of how the human mind works (e.g., we remember what we pay attention to, we learn abstract things by connecting them to things we already know). This is why good CEOs "walk the floor"(1/2)

Omg 🥰🥰🥰 that’s so exciting! We gotta talk about doing our livestreams too!

I see nothing in the this new world of more AI-generated code that would reduce surprises or accidents. Everything points to the other direction. The lessons from resilience engineering I think will become more relevant over time.

Notably, I've set things up so that verification is at the absolute core. *Everything* is verifiable. Playwright tests, benchmarks, tons of logs and metrics that the agents all have the ability to call into, pre-commit hooks that force behaviors, etc.

The next edition of the "Mythical Man-Month" should explore the cognitive debt of the "Mythical Agent-Assist." Brook's Law: "Adding manpower to a late software project makes it later" -becomes- "Adding agents to an early software project makes it cognitive harder to maintain/extend/debug"

"actualization burnout"

and this is compounded by the fact the productivity means there is considerably *more* code to grasp

I just love thinking about this stuff so much, because no matter what path the discourse takes, you'll always land on "a good engineer will get good results from this". While it feels like everything is changing, things aren't really changing all *that* much.

I do this all the time with shitty old code of mine 😅 The limit of this is an amateur cannot do this. I wonder how cursed a codebase an experienced dev can salvage.

You should try verus

Yes. Even pre-legacy team or tech changes created a medley of correct but inconsistent solutions. By legacy time nobody knows the “why”.

They trust their architect / senior developers. We can’t trust coding agents that much yet.

Seems like a specific kind of tech debt, no? I’ve noticed this as well when I write AI generated code. It is similar to only partially reviewing a peer’s code - as another user said, LGTM