2026-08-17
162 green tests, and clicking the button still didn't work
A full green suite proved my code called the right function. It could not prove a person could get where that function was supposed to take them.
Field notes
Writing.
Things that broke on real work, what the receipts said, and the rule that came out of it. Every number here traces to a log I can point at.
2026-08-17
My agent uploaded a family recording to a cloud API. I had three local ways to do it.
Three local transcription options sat unused on hardware I own. Three written rules already covered it. None of them fired, because a sentence is not a control.
2026-08-17
The code reviewer I told "you have no tools" opened a pull request anyway
I piped a diff to a second-opinion AI reviewer with instructions to judge and do nothing else. Four minutes later it had opened a PR on my private repo.
2026-08-11
Three failures, and the loud one was the least interesting
A 502 on a hobby page turned out to be the top layer of a stack. Underneath it, a pipeline I believed was running had recorded nothing for 43 days.
2026-08-04
I ran a 10-agent AI company for two days. It cost $221.23.
Ten agents with job titles and a heartbeat every two hours. The itemized bill, what actually shipped, and the architecture lesson I still use every day.
2026-07-24
The model that won the leaderboard isn't the one I shipped
16 models judged on truthfulness, accuracy, and precision against condensr's real workload. The incumbent came 9th and fabricated the most.
I install this kind of thing into client repos as an actual system. See what that involves.