Writing
Notes on agent design, evaluation, and the failure modes that only show up in production. Mostly things learned the expensive way.
Own the memory, rent the model
Sep 2026Five months running Karpathy's llm-wiki pattern: what a portable memory layer actually buys, where it's worth diverging from the gist, and the three ways maintenance quietly fails.
Write the standards, not the steps
Sep 2026An autonomous research agent had to move to a domain it was never built for. The port took two days — and the reason was a ratio: nine parts standards, one part domain.