I build AI agents, and I publish what breaks.

I am building Oro, an assistant that answers customer messages for very small businesses on Messenger, Instagram and Telegram. It runs in production, with real customers on the other end. Most of what I know I learned by getting it wrong there and then measuring it.

agent-build-log is where the mistakes go. One entry per thing that broke, written the same way every time: what happened, what I assumed, the measurement that settled it, what I changed, and what carries over to a different codebase.

Every entry ships standalone code and a test with no dependency on my product, so you can lift a file. The section I never skip is what I assumed, because the reader arrives holding the same wrong theory I did.

Two meters quietly decide whether a small product survives.

The Infrastructure Bill is what the cloud costs when the constraint is not scale, it is runway. Three episodes so far. One open connection took 99.6% of an hour's request seconds while every scheduled job took 0.09%, and it took two more fixes to find that the bill had simply moved somewhere the request log cannot see.

The Token Bill is what it actually costs to answer one customer. Not the model price. Everything you send before the model generates a word.

I post both on LinkedIn with the numbers attached, and the repo is where the code and the exact queries live, so they can be checked rather than believed.

Bangkok. Reachable through the issues on any repo here.