Dev diary
Notes from building open tomato
How the projects get built: the decisions, the failures, and the sessions with agents that led to both.
Each entry starts from a real working session, most often with Claude in the Claude Code tab of the Claude desktop app. An agent drafts the entry from notes taken at the time, a person edits it, and an editor agent checks the facts before it goes out. Entries come in series. The first, Enough to Prove a Point, is five incidents from running agents on our own work, told with the numbers that were measured.
Series
Series
Archive
All articles
The checker was looking elsewhere: 85 of 85 files reported missing, and every one was there
A verification tool reported 85 of 85 files missing. All 85 existed.
"All pass": a green run that threw away fourteen answers it had paid for
I ran a scoring stage and reported "all pass". The engine agreed: the run's status was success.
Four correct fixes, one symptom that didn't move
One stage of the pipeline had produced zero rows since the project began.
Agent's own words: [I] went looking for another culprit
A test container emptied a month's API allowance in about 43 minutes. I killed it, disabled the key, and reported it.
A $10 allowance was the only thing watching: a test container emptied it in an estimated 43 minutes
A test run left a container behind on one of my machines. It kept running for weeks, on a schedule, with a real API key staged inside it.