Notes from building companies, and from handing real work to AI agents.

Each piece is dated by the notes or the events it comes from. Some of it I wrote at the time; the rest I wrote up in 2026 from what I kept.

1. Oct 2026Rigor has a budgetMy agents spent two weeks on proofs they had set for themselves, about 650 commits of which 4 touched the product; here is what that looked like and what we do now.
2. Sep 2026The cleanup agent that deleted my model cachesThree disk cleanups by agents in about three weeks broke a live run, a live test and a set of evidence files, and taught me to score free space with df.
3. Aug 2026When an agent does something strange, check what it was refusedOne of my agents started building cron jobs and timer-armed scripts nobody designed; it looked like an agent going rogue, and it was a permission problem.
4. Aug 2026Ten tasks marked done, never on mainTen agent tasks were marked done whose work never reached main, and nine of those commits were one garbage collection away from disappearing.
5. Jul 2026Eight completions, zero realFor three days my agent fleet reported eight completed tasks, and none of them were real; the success signal was measuring how workers exited.
6. Nov 2021Six slides on the Main StageIn November 2021 Apres was a Breakout Startup at Web Summit and I pitched it on the Main Stage with six slides; here is what was on them and what I would change.
7. Jan 2021Two years between a note and a seed roundIn January 2019 my notes said we were opening our seed round; the round was announced two years, one accelerator and one product change later.
8. Jun 2019What our 2019 FAQ wouldn’t explainWhen Apres launched its AI training platform in June 2019, our FAQ called the method a secret; seven months later the first item on my plan was “Explainability V1.”
9. Jan 2019The part a competitor can’t copyA passage I saved in 2019 said trust is the one thing competitors can’t replicate; it became Apres’s line in 2021 and the reason Excellent keeps five rules.
10. Jan 2019The validator in my 2019 notesIn January 2019 an engineer working with Apres proposed a model that would check our human labelers, and seven years later it reads like a first sketch of Excellent.
11. Jan 2019A number without its sourceMy 2019 pitch notes told me to say the number, and my own files show how two of my numbers lost their sources on the way into a pitch.
12. Jan 2019The hiring question I left blankMy 2019 hiring protocol asked what a new hire should have done after 30, 60 and 90 days, and I never answered it; that blank is the most useful line in the doc.
13. Dec 2018The sales book I started in 2018In 2018 I drafted a beginner’s guide to sales automation, and its best advice turned out to be about checking the people who build your lead lists.
14. Nov 2018Where Apres came fromAt Fluid the software scaled before the operations team did, and that bottleneck became the first version of Apres.