Notes from the work

Writing

Notes on building software with coding agents and language models, from my own projects.

  1. Garden plans that nobody checks

    A model designs the garden and code decides whether the plan goes out. No person reads it first.

  2. Two days of choosing a chat model for Polish

    Google gave the model behind CityChat's chat an earliest shutdown date. Switching to the suggested replacement would have been one line of config; I measured first, and on the second day the answer changed.

  3. Where my rules for coding agents came from

    CityChat has merged 981 pull requests since June, some of them written by agent sessions. The rules those sessions work under fit on one page, and most of them came from an incident.

  4. Fact-checking a database with research agents

    Research agents are good at finding sources and bad at quoting them. What I check after them, learned on a run over about 14,900 places.