Notes

Working notes on building software with AI agents doing a large share of the work. The subject is the system around the model rather than the model itself, and the code underneath both: how sessions coordinate, where the rules live, and which checks can actually fail. A case-study series runs alongside them and is still being added to, so it sits apart from the dated notes.

  1. Latest

    Holding it right

    One talk made a weak model do a real job with a harness. The other says no harness fixes what agents do to a codebase. I stopped reading every line, so I checked where my own checks stop.

  2. One line, fourteen packages

    One dependency line pulled 14 packages and about 40,000 lines of code into the app. Apple closed most of the gaps and an agent writes most of the boilerplate now, so what does a third-party architecture still buy?

  3. One caller hid the crash

    A test suite passed seven out of seven while the code under it corrupted the task allocator on every run. Only one file called the initialiser that triggers it, and the bug needs two.

  4. Asking if it's finished strips the lock

    The polite way to coordinate two AI sessions sharing one save file, asking whether it has finished, is the message that strips its protection. I caused the bug live while researching it.

  5. A correction looks like a duplicate

    A file my agents keep said the opposite of the truth about my own machine for weeks. A June paper measures why corrections get thrown away by the check meant to stop duplicates.

  6. The template wrote the ticket

    An agent filed nine paragraphs to change two words in a comment. It had followed its instructions exactly.

  7. Coordination lives in the harness, not the prompt

    A session hands itself off to the next one, and won’t trample work another session is already doing. Neither behaviour is a prompt.

  8. Your rules should travel with you

    The rules you build with an AI assistant are professional capital. They belong to the studio, not the vendor.

  9. Notes on coordinating AI agent swarms

    The primitives missing from agent-swarm runtimes, and the disciplines that fill the gap.

  10. Thirty days of agentic development

    I counted what thirty days with an AI coworker actually produced, and about a third of it never reached git.