Here's what I'm missing in all the TDD Agent Harnesses I've seen. No one's asking the question: Is the existing design ready for the next red test? Or do we have to change something first?
The Agent is just crushing through like a bulldozer, no questions asked.
Well, that's not how this works.
Programmieren war nie “eine Sprache sprechen”. Es war: einen Gedanken zu Ende denken bevor man ihn ausspricht. Syntax war nur die Verpackung. Das Produkt war immer der Gedanke.
It’s easier to see a problem than to avoid it.
Even for LLMs.
Models can often explain what’s wrong with an answer more reliably than they can avoid generating the flaw in the first place.
Detection is a classification task.
Avoidance is a generation task.
Complexity Wall
(noun)
The point where a system becomes so complex that there is no intelligence left to make progress in a reasonable time.
Agentic coding won't work with your PR workflow. It will incentivize larger, crappier, unreviewable, LGTM batches to hit the complexity wall fast.
We need is a workflow that enables MMMSSMF (Many More, Much Smaller Steps, More Frequently), so that we can speed up while still being able to review it.
Spec-first approach is counterproductive when using AI. What AI brings to the table is superpowers to engage in serious exploratory, experimental approach to engineering.
In agentic coding, don't obsess over rules.
Build tight feedback loops and iterate.
You gain far more from iteration than from rules.
Join us for a day of coding excellence -> Global Day of Coderetreat 2025 in Vienna -> insights.squer.io/gdcr-2025 #gdcr #coderetreat
The richest man on earth owns X.
The second richest man on earth is about to be a major owner of TikTok.
The third richest man owns Facebook, Instagram, and WhatsApp.
The fourth richest man owns The Washington Post.
See the problem here?
Given that we now understand we’ve been training AI to guess rather than to admit when it does not know, it should make us reflect on exams and the way we educate people.
We never needed a senior doing the work of 10 juniors.
What we need is a senior doing the work of a senior, together with a junior.
I tell my agents to conclude with their confidence level.
GPT-5 almost never gives me [10/10 confidence], while Claude Sonnet 4 does it very often.
To think higher confidence is better would be foolish.
Elon Musk spoke by video to Tommy Robinson's anti-immigrant rally in the UK today.
"You're in a fundamental situation here where, whether you choose violence or not, violence is coming to you," said Musk. "You either fight back or you die."
Russia violates NATO airspace, and Trump acts like he’s live-blogging The Bachelor.
No other POTUS in history would make such flippant posts and statements about attacking an American city. The fact that this stuff is commonplace in the current administration and tolerated by so many is testimony to the moral rot we are suffering as a nation.
Considering the effort I spend making sure the agent only reads what it has to and to bring this into the simplest form, eliminating all noise, and to ration this in small as can be contexts.
It makes me wonder.
Why didn't we do those things earlier - for ourselves? Do we not care?
A lot of people were asking about examples for my "Augmented Coding - A Pattern Language" post. So I was iterating over it and added some. I also added new patterns.
gregorriegler.com/2025/07/12/a...
How is it that Atlassian builds tools that:
- SO MANY people use
- SO MANY people dislike (JIRA, Confluence. Previously also HipChat)
They are defying the conventional wisdom for business success that starts with "build something people will love"
Paris hat die letzten Monate für 9,6 Mio. Euro seinen Rathausvorplatz entsiegelt und einfach in einen neuen Stadtwald verwandelt. Das Ergebnis sieht so wunderschön aus❤️
As @socrates-conference.at 2025 is getting closer: Which sessions did you enjoy most last year? Quick reminder: We had all these socrates-conference.at/schedule/
For me, it were the #DDD discussions and the "Coding Fun" sessions with @codecopkofler.bsky.social
🔬 Real Browser Behavior Modern browsers like Chrome, Firefox, Safari, and Edge all follow this exact logic — because they conform to the official CSS standard. If you try this: html Copy Edit <style> #box { color: red; } /* (1, 0, 0) */ .a.b.c.d.e.f.g.h.i.j.k { color: blue; } /* (0, 11, 0) */ </style> <div id="box" class="a b c d e f g h i j k">Hello</div> The result will be blue text, because the rule with 11 classes has higher specificity.
A problem with modern LLMs is that they are so convincingly anthropomorphic now that when they produce wildly incorrect responses like this one, my reaction isn't "oh, I found a bug", it's "you're a lying sack of shit!"... and that level of emotional reaction to a piece of code really isn't healthy.
I expect that consumer-facing AI programs will continue to improve and they may become much more useful tools for everyday life in the future.
But I think it was a disastrous mistake that today’s models were taught to be convincing before they were taught to be right.
When you're faced with a problem, don't think: "What's a possible solution?"
Think: "What information am I missing?"
I had a lot of fun doing augmented coding as of recent and decided to capture my learnings in a new blogpost.
Here is what I learned:
Augmented Coding - A Pattern Language
gregorriegler.com/2025/07/12/a...
Pro Augmented Coding Tips:
-Define the process using .md
-Keep context small
-The agent creates it's next context
-Preserve cross-context memory in .md
-Keep the code running
-No new code without a failing test
-Filter needless script output
-Small steps and iterate
-Iterate on your process files
I am trying to Roll Forward when I should have Mikado'd.
We all need reminders to take smaller steps from time to time.
The Efficiency Delusion
What do you mean? Like you select only a part but it still extracts the whole thing?
