ORION LEE · WRITING
TechnologyBlog
Systems, software, AI, and the craft of making complex things dependable.
0102
The Agent Said “Done.” Did the System Actually Change?
A timeout can hide a successful write. Designing reliable AI agents means preserving intent, reconciling uncertain outcomes, and requiring evidence before declaring success.
Prompt Optimization Is a Living Evaluation System
Start with Product and QA good cases, evaluate real outputs continuously, and let dimensions and thresholds evolve without pretending everything is measurable.