Notes from building with AI agents
How I learned, what broke, and how each fix was proven. Written from my own work notes, with every employer detail taken out.
How I got here
- From zero to a production platformHow someone trained in physiotherapy learned to build software with AI, one written-down failure at a time.
- Building a platform on a locked-down desktopNo admin rights, no database server, no open ports. How the order platform's architecture came out of its constraints.
Problems and proofs
- Measure the theory before building the fixEveryone suspected the file server. A canary and a few counters pointed at our own service, and the fix worked on the first try.
- The ERP as the answer keyNobody had time to label data. The filed paperwork already paired every purchase order with the order the ERP booked. Whole-order accuracy went from 18% to 64% in nine days.
- One action, one control, one logicA fix added a second way to put a file into an order. I made a rule that a second way is a defect, and the agents proved it door by door.
- “Not released” can be the gate workingA small, well-proven change, a forced logoff, and 44 minutes against a 95-minute test run. The agent reported not released. That was the right answer.
- Prove it as every roleA supervisor's order board hung on “Loading”. Every proof had run as an administrator.
- Measure the new machine before touching itThe workstation I built on was shut down at a few hours' notice. Two of my own claims about its replacement were wrong within a day, so a probe measured it instead.