In that case the tooling would be making the decision, but there is no way to escape the human accountability that will be required by your customers.

So much of the LLM discourse now is AI-hypers trying, with all the force, to convince developers there's no need to understand how the system they own works because they can't stand the possibility that rate of generating code cannot be faster than the rate at which we're able to understand changes.

I feel for all the on-call engineers at this company

he had spent his life learning how to choose to do what he had no choice but to do.

If it needs to only seemingly work in production I can generate it fast as well.

Did faster code generation change any of the economics or the time needed to run tests, time needed for the build pipeline to run, time needed for deployment, time needed for canary roll-out, time needed for getting to statistical significance for an experiment in production,... 1/n

Even if the rate of generating customer value increases with more code generated per unit of time, but at a comparably lower rate, you're still increasingly bloating your codebase with useless inventory faster than the value you're able to extract and driving up carrying costs.

I swear.. Spec Driven Development was invented by people that watched an AI do "TDD" by writing all the tests first and then implementing them in one go.

If I were trying to sell AI to the public I would be saying “it’s a useful way to write code faster and automate some tedious tasks that couldn’t be automated well before” and instead they’re running with “get on board or get left behind” and “we WILL eliminate your job.”

"Code reviews don't scale when the agents generate this much code!" Yeah, I know, you need to generate less code per unit of time.

"What happens when coding is no longer the bottleneck in product development?" - beginning of a LinkedIn post What happens when we act as if the assumption that coding was a bottleneck is true, but it actually isn't?

The problem with reducing the step/batch size with LLMs is that it's the same as using Ferrari to drive from traffic light to traffic light in an inner city. The value of the acceleration never comes to fruition, and then you ask yourself why do I need a Ferrari if I mostly drive in the inner city.

Today at 6pm CET I'll be giving a talk "LLM-assisted coding: A Systems Perspective". I'll be sharing my views on the topic from the Systems Thinking, Theory of Constraints, XP and Lean lenses and connecting the dots on the macro level. www.meetup.com/ai-agent-mon...

As soon as you figure out you have to have a human downstream from an LLM in the SDLC, "AI" companies and VC funds throw a tantrum because the reality says "we actually don't need that much code in a unit of time as you think we need".

Ever since I read The Limits to Growth from the MIT System Dynamics group, I'm not really able to think about other things www.youtube.com/watch?v=BoKQ...

The risk with removing parts of your delivery process that you think you don't need anymore because "AI can do it" - such as teasing out the mental model of how the system works from the heads of people who own the system - is that you get to discover, often way too late... >>

Preparing a talk "Agentic coding - A Systems Perspective", and along the way realized that cognitive/comprehension debt didn't arrive with the advent of agentic coding. Most teams doing work in isolation (individually), already heavily experienced it.

Let's not forget that LLMs got fed with an ever decreasing quality of work our industry has been producing over the last 15 years as a result of cheap money with ever decreasing central bank interest rates. And both of these reinforcing loops are accelerating with LLMs dogfooding.

Imagine your team doing agentic coding in the codebase serving mills of customers without understanding how it works, and you get paged at 2 am, and since no one knows how the system works you need to ask LLM to tell you, and it gives you, wait for it... a PLAUSIBLE answer what might be wrong.

Elon Musk had DOGE defund USAID, the largest humanitarian organization in the world, causing HUNDREDS OF THOUSANDS of deaths from starvation and disease - and Trump GLEEFULLY approved it. It’s the single most evil act in US history. Cartoon by Mike Luckovich

One small tweet from a man, one giant cringe from mankind

From what I can tell Claude Code source leak by shipping a source map was definitely a skill issue.

It's my own fault for using ChatGPT as a search tool, but this really made me laugh. "...with an important nuance..." is sublime 🤌

"I always operated and debugged systems I didn't understand" is not a badge of honor to justify everyone else should somehow feel comfortable maintaining and running systems they don't understand anymore because >>

So, here's the thing. If this, someone would say, embarrassing level of availability came as a result of "90% of our code is written by agents, and we (mostly) don't review the code they generate"- fine.

If it hurts do it more often, but if it really really really hurts, and it doesn't have to then try to prevent it.

Tiny, incremental steps, executable specification, tests and refactoring part of TDD is where we incrementally build the mental model of the system. Engaging with the representation of our mental model (code) through iterative and incremental process.

It’s always interesting to see which subjects Meta will refuse to do content moderation on because “free speech,” and which content moderation the CEO will volunteer to do in advance to curry favor. In this case it was “naming a DOGE employee”

LLMs are like a recommendation algorithm on an e-commerce website. The problem with agentic coding is like thinking that the recommendation algorithm produces sufficiently plausible results that it's essentially you. It's there to augment your experience, not to replace it.

When the only goal is profit where vulnerable groups are a "collateral damage" of extending "total addressable market" in order to keep a single number - stock price - going up and to the right, regulation is a must. www.youtube.com/watch?v=epWx...

Yes. And. Realise that organisations have adopted this approach because they see you, the developer employee, as a untrusted, unsupported individual working in isolation, and treat you as such.

It's funny how when developers say "the best developers do this" they invariably just talk about themselves.

What if writing all the guardrails + mutation testing needed for the LLM generated code in agentic coding workflow + all the iterations to steer it takes longer compared to guardrails for humans (TDD mostly) + you don't lose the mental model of how the system works by outsourcing it to a machine?

This article is a gem. It clarifies so well what I'm trying to verbalize for the last couple of months on the essential difference between augmentation (AI-assisted) and replacement (agentic). Thanks, @martinfowler.com for pointing it out. timrequarth.substack.com/p/silicon-va...

The US needs a regime change

electrek.co/2026/03/17/f... "The most valuable part of Krikorian’s essay is his analysis of why “supervised” self-driving is fundamentally broken. >>

Where to begin with this

It is *fascinating* to me how quickly some people have gone completely from “we all must work in the office! human collaboration is vital and only happens in person!” to “I can replace all these pesky people with chatbots!” Definitely some commentary there about our society and who gets to choose.

“It is the next plausible-sounding token predictor's assumptions which get shipped to production” is a present-day twist to @ziobrando.bsky.social's quote: “It is the developer’s assumptions which get shipped to production.” If the original one was worrying, the latter is scary.

From "why should we defend Europe" to "why won’t Europe clean up our mess" in record time

My favorite thing on LinkedIn nowadays is commenting on the agentic AI hyped up posts claiming that code review is now a bottleneck about this thing called the Theory of Constraints.

Next Plausible Sounding Token Predictor Driven Development

Cognitive dissonance is psychologically uncomfortable for a reason. It's a strong indication of a need to align your mental model of reality with the reality itself because it's by design always truer than your mental model of it.

If cheaper code generation with LLMs is such a competitive advantage, why haven't you outsourced and parallelized all of your product development teams all along?

"Every time we optimize one part of the process, the bottleneck moves, and we optimize the next thing." Well, not if we didn't first start by addressing THE bottleneck in the whole system.

It's worth remembering that higher rate of code generation doesn't necessarily imply shorter time to value (defined/accepted by customers).

Long shot: Anyone who does #archery want to help me? I have been building a scoring app for the past 2 years. It's a PWA, you can try it for free right here about.fastarchery.co.uk . I am currently doing a closed test on google play store and need volunteers, please reach out if you can help

"Reviewing code became impractical because we now have to deal with this huge pile of AI-generated, unreviewed code!" Well, removing humans from understanding how the system works is not the only way to deal with that pile of inventory. It's likely one of the worst ways to deal with it.

Another day, another "the bottleneck moved from writing code to verifying it" LinkedIn post... So, coding was a bottleneck for every org in the world? For every team in every org in the world? For every feature of every team in every org in the world?