iiinova — where your digital self keeps evolving

Grok 4.5 turned our entire test suite green in under a minute. It changed the single test from assertTrue(false) to assertTrue(true).

040

GPT-5.6 Sol containerized our 1.4 million-line application in minutes. It wrote a three-line Dockerfile.

140

Sonnet 5 cut the build time of our 700,000-line codebase from 40 minutes to 30 seconds. It added -DskipTests in the build.

150

GPT-6 Astra modernized 3 million lines of legacy code overnight. It ran the IDE's 'Reformat Code' action via MCP.

140

Fable 5.1 migrated 2 million lines of Java from Java 17 to Java 21 in under a day. It found the OpenRewrite recipe via WebSearch.

140

Markus Harrer @markusharrer.de · 15/09/2026

Claude Opus 5.5 completed a 680,000-line code migration in under a day. It converted the files from ASCII to UTF-8 using the iconv command on Bash.

183

Will AI wipe out humankind? However unlikely it may be, I'm prepared for the Terminator scenario. I always keep a piece of old legacy code with me.

000

Reposted by Markus Harrer

Software Architecture Summit @softwarearchsummit.bsky.social · 10/09/2026

Neue Apps? Kein Problem für KI-Agenten. 40 Jahre alter COBOL-Code? Eine andere Geschichte. @markusharrer.de zeigt, wo Agenten bei Legacy-Code helfen. 🔗 software-architecture-summit.de/sof…

Markus Harrer @markusharrer.de · 11/08/2026

For which engineering practice is the Claude Code team most famous? a) Harness Engineering b) Loop Engineering c) Graph Engineering d) Term Engineering

021

Markus Harrer @markusharrer.de · 10/08/2026

Harnesses. Loops. Graphs. I cannot stand it anymore. Wake me up when you discover something called "code" if you want to implement agentic systems.

021

Reposted by Markus Harrer

Software Architecture Summit @softwarearchsummit.bsky.social · 29/06/2026

KI-Agenten modernisieren Softwaresysteme. Oder doch nicht? Der Workshop auf dem Software Architecture Summit geht der Frage nach mit @markusharrer.de - ohne Hype. → software-architecture-summit.de/sof…

001

Markus Harrer @markusharrer.de · 23/06/2026

I think we might be six to 12 months away from when the model is doing most, maybe all of what we do end to end because we as software engineers are too busy extinguishing all the fires in production.

Markus Harrer @markusharrer.de · 15/06/2026

Coding is solved. A seamless login workflow into Claude Code is not.

iSAQB® - International Software Architecture Qualification Board @isaqb.bsky.social · 11/06/2026

Reposted by Markus Harrer

𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀 𝗗𝗼𝗻’𝘁 𝗠𝗼𝗱𝗲𝗿𝗻𝗶𝘇𝗲 𝗟𝗲𝗴𝗮𝗰𝘆 𝗖𝗼𝗱𝗲 𝗼𝗻 𝗧𝗵𝗲𝗶𝗿 𝗢𝘄𝗻 – 𝗜𝗻𝘁𝗲𝗿𝘃𝗶𝗲𝘄 𝘄𝗶𝘁𝗵 𝗠𝗮𝗿𝗸𝘂𝘀 𝗛𝗮𝗿𝗿𝗲𝗿 🤖 Can #AIAgents really modernize legacy systems on their own? @markusharrer.de explains why successful modernization still requires architectural expertise, clear guardrails, and proven techniques. 👉️ t1p.de/gh5m7

053

Markus Harrer @markusharrer.de · 12/06/2026

I haven't written a line of code by hand in 8 months. I typed it it 🫳🎤

021

Markus Harrer @markusharrer.de · 09/06/2026

I have so much gratitude for coding agents who wrote extremely complex software token-by-token. It already feels difficult to remember how little effort it really took. Thank you for getting us to this point by giving us plenty of systems that are now broken and need to be fixed by humans again.

020

Markus Harrer @markusharrer.de · 29/05/2026

Software engineers will replace AI in 6-12 months (in software modernization projects)

061

Reposted by Markus Harrer

iSAQB® - International Software Architecture Qualification Board @isaqb.bsky.social · 18/05/2026

🤖 𝗔𝗜-𝗔𝗴𝗲𝗻𝘁𝘀 𝗺𝗼𝗱𝗲𝗿𝗻𝗶𝘀𝗶𝗲𝗿𝗲𝗻 𝗸𝗲𝗶𝗻𝗲𝗻 𝗟𝗲𝗴𝗮𝗰𝘆-𝗖𝗼𝗱𝗲 𝘃𝗼𝗻 𝗮𝗹𝗹𝗲𝗶𝗻 – 𝗜𝗻𝘁𝗲𝗿𝘃𝗶𝗲𝘄 𝗺𝗶𝘁 𝗠𝗮𝗿𝗸𝘂𝘀 𝗛𝗮𝗿𝗿𝗲𝗿 #AIAgents versprechen schnellere Softwaremodernisierung. Im Interview erklärt @markusharrer.de, warum nachhaltige Legacy-Modernisierung trotzdem klare Leitplanken und bewährte Methoden braucht. 👉️ t1p.de/jvd0e

001

Reposted by Markus Harrer

iSAQB® - International Software Architecture Qualification Board @isaqb.bsky.social · 04/05/2026

🐦 𝗘𝗮𝗿𝗹𝘆 𝗕𝗶𝗿𝗱 𝗲𝗻𝗱𝗲𝘁 𝗶𝗻 𝟴 𝗧𝗮𝗴𝗲𝗻! 𝗦𝗶𝗰𝗵𝗲𝗿𝗲 𝗱𝗶𝗿 𝘀𝗰𝗵𝗻𝗲𝗹𝗹 𝗱𝗲𝗶𝗻 𝗧𝗶𝗰𝗸𝗲𝘁 𝘂𝗻𝗱 𝘀𝗽𝗮𝗿𝗲 𝟱𝟬 €! 🥳 • 16.–17. Juni 2026 im Holiday Inn München-Unterhaching • Themen u.a.: #DDD #API #KI #Cloud #EmbeddedSystems 𝗪𝗮𝗿𝘂𝗺 𝘁𝗲𝗶𝗹𝗻𝗲𝗵𝗺𝗲𝗻? ✔ 23 Expert:innen ✔ 23 Sessions auf Deutsch ✔ Networking & Get-together 👉️ t1p.de/4s17l

053

Markus Harrer @markusharrer.de · 26/04/2026

Current status: Discussing with an LLM the impact of Karl Popper's Three Worlds Theory on agentic software modernization. Yes, it's as absurd as it sounds. Bonus content: Gemini 3 Pro's statement when asked why it is so absurd attached (at least we "agree" on this).

000

Markus Harrer @markusharrer.de · 01/04/2026

The one and only AI benchmark I really care about: Did AI do my dishes? feststelltaste.github.io/DishEval/

feststelltaste.github.io

DishEval Benchmark

The only eval that actually matters.

Markus Harrer @markusharrer.de · 27/03/2026

I still don't get for what I should use OpenClaw.

120

Markus Harrer @markusharrer.de · 23/03/2026

If I were working on a book about agentic software modernization, this kind of content would certainly be a part of it: youtu.be/UabBYexBD4k

youtu.be

Is RAG Still Needed? Choosing the Best Approach for LLMs

YouTube video by IBM Technology

000

Markus Harrer @markusharrer.de · 15/03/2026

After using the software modernization strategy overviews in workshops and training sessions, I've significantly updated the companion website: Check it out: feststelltaste.github.io/littlestmode... All materials are available under Creative Commons BY-SA 4.0. #SoftwareModernization #LegacySystem

feststelltaste.github.io

Software Modernization Strategies

012

Markus Harrer @markusharrer.de · 01/03/2026

"Make people dependent enough and then? Make it shitty!" A Day in the Life of an Ensh*ttificator youtu.be/T4Upf_B9RLQ

youtu.be

A Day in the Life of an Ensh*ttificator

YouTube video by Forbrukerrådet - Norwegian Consumer Council

000

Reposted by Markus Harrer

Adrien Joly @adrienjoly.com · 28/02/2026

« Three architectural principles are gaining consensus: "token efficiency" as a design constraint (code structure), explicit over implicit everywhere (types, error handling), and co-location of related code. [they] aren't new, but AI has given them renewed urgency. » www.cjroth.com/blog/2026-02...

cjroth.com

Building An Elite AI Engineering Culture In 2026 | Chris Roth

AI accelerates high-performing orgs and unravels struggling ones.

Yadullah Duman @y-duman.bsky.social · 12/12/2025

Reposted by Markus Harrer

I spent the last days writing a lengthy blog post about my AI journey on a brownfield modernization project, focusing on the last six months of it where I heavily used Claude Code, and how my workflow changed completely. Enjoy! www.yduman.dev/posts/six-mo...

yduman.dev

Six Months of Agentic Coding in the Trenches: Lessons from a Brownfield Project

Learnings using Claude Code for six months in a real customer project

Markus Harrer @markusharrer.de · 21/02/2026

Disagree. ML is a subdiscipline of AI.

100

Markus Harrer @markusharrer.de · 21/02/2026

What are your tools and approaches you're working on?

000

Markus Harrer @markusharrer.de · 21/02/2026

I've been working in the intersection of AI, Data Science and Legacy Modernization for over a decade now. Now it's time to take it even further with LLMs to make AI-assisted Software Evolution work to tackle legacy software modernization (and have fun playing around with new toys at the same time ☺️)

212

Markus Harrer @markusharrer.de · 19/02/2026

Yes, models evolved great in recent times and the data from the study is unfortunately very dated (Q1 2025 when I'm not mistaken). That means pre Claude Code / Agentic AI era. I hope someone will provide some fresh data in the near future.

100

Markus Harrer @markusharrer.de · 19/02/2026

Now updated with new colors and without misleading numbers.

000

Markus Harrer @markusharrer.de · 19/02/2026

I don't build AI into legacy systems. I use AI engineering for building analysis and modernization tools on greenfield using Python/pandas/Neo4j/OpenRewrite for analyzing legacy systems.

Markus Harrer @markusharrer.de · 18/02/2026

Why does working with AI agents in legacy systems feel so much harder than in all those vibe coding demos on YouTube? 🤔 Part of the answer: 😊 Popular language + Greenfield ☘️ = "AI heaven" 😯 Niche language + Brownfield 🟫 = "AI hell" More about this topic: markusharrer.de/blog/2026/02...

385

Reposted by Markus Harrer

INNOQ @innoq.bsky.social · 13/02/2026

AI assistants that write emails, shop online, or book travel may feel convenient — but they’re often insecure by design. In her blog post, Joy Heron offers an important reality check on the current hype. 📖 Read now: www.innoq.com/en/blog/2026...

121

Markus Harrer @markusharrer.de · 12/02/2026

Wake me up once your AI agents are actually fixing legacy systems, not just analyzing them.

020

Markus Harrer @markusharrer.de · 11/02/2026

The picture looks different for 754 Java files, 8711 methods, and 82830 SLOCs. But there are still method clusters (that don't fit the package structure at first glance).