02Writing

What breaks when an agent meets the real world.

Patterns and lessons from building an autonomous testing agent, the computer vision behind it, and what running it in front of real customers costs. About one post a month.

• Sept 2026

  CoTester: building an LLM agent that checks its own work

  Why generated steps break, how self-validation catches it, and what nine versions in front of real customers actually taught.

• Sept 2026

  Cutting cost per conversation in a production LLM agent

  Telemetry worth collecting, where model selection actually pays, and what to measure before optimising anything.

• Sept 2026

  Finding UI elements when the DOM won't cooperate

  Transfer learning on a small hand-labelled dataset, and where a vision model beats a selector.

• Sept 2026

  What breaks when an LLM agent meets real users

  A field guide to the failure modes that only appear once people who did not build the agent start relying on it, one per section, with what worked against each.

New posts land here about once a month. To be told when they do,subscribe by RSS: paste the link into any feed reader.