How I Keep Thinking Critically When AI Agents Do the Work
The better the agents get, the less anything forces me to think. These are the verification rules I built into my daily work so it does not depend on my mood.
23 posts on this topic.
The better the agents get, the less anything forces me to think. These are the verification rules I built into my daily work so it does not depend on my mood.
KI Challenge 2026: StatAgent computes a monthly employment statistic without AI, built in under 21 hours with coding agents. Process, numbers, lessons.
Double-entry bookkeeping for businesses that file a balance sheet: BuchhaltGenie Doppik, planned for 1 December. Why a second product, and what is missing.
One page pair needs registering in five places. A hand-kept list silently drops one of them. So I derive all five from a single typed registry.
A RAM warning fired on 84 percent of all session starts with zero OOM events. The fix was never the threshold, it was measuring the right quantity.
22 of 30 sessions run zero waves. The fast path built for that case writes zero events and is unlogged in 258 of 265 runs. A recipe for finding yours.
A render call could take 46 seconds inside a webhook with a 30 second budget. The second finding turned out to be worse than the first one.
One week after HACK_002: a real bug in the court deadline is fixed. New are a deadline calculator without AI, advice centres with contacts, and Turkish.
A script found the same work under eleven job titles. It turned into BewerbungsGenie, open as a public beta for Austria since 1 October 2026.
A family photo album made with coding agents: from a dictated prompt on Thursday evening to the first production deploy on Friday, and what is still missing.
HACK_002 recap from Vienna: the Pflegegeld-Prüfer, a care allowance checker, wins Track A1 but misses my advance target of 12 of 14 court cases. It gets 10.
Kleinunternehmer spans three laws in Austria. Why V2 kept exactly one column and reads the threshold from a date-gated value table.
Two hackathons in Vienna, one week apart, and the building only starts at the event. What I still get ready beforehand, and why my code stays home until then.
A bakeoff across 16 real receipts exposed a systematic factor-1000 error. Why a deterministic guard sits above the model, not inside it.
About 20 people paid for my product, not one of them twice. What the numbers actually say and why the gap sits in activation, not in acquisition.
One /discovery run, several parallel agents, a live-browser test: how I had my own website audited, and why a human still makes every final call.
My agents leave zombie processes behind that quietly eat my RAM. So I built myself a small macOS menu bar app that hunts them down and kills them off for me.
Three Vienna appearances in two weeks: a career panel, a live session-orchestrator walkthrough, and a software architecture panel.
A Whisper default replaced 13 minutes of my first real end-to-end transcript with a hallucination loop. The diagnosis, the fix and four transferable lessons.
AI-assisted development has shifted its bottleneck from writing code to judging it. Why verification now matters, from daily practice, not theory.
Qwen3.6-35B technically loaded on an M4 Pro 24 GB at 100% GPU, yet ran at 0.1 tokens per second. A quiet failure, five lessons and hardware recommendations.
A harness gives AI coding agents context and verification rules: CLAUDE.md, Rules, handovers and quality gates for long-running projects.
What sits between an AI prototype and daily use: data, failure handling, ownership, and deciding when an experiment has answered its question.