When I was building a skill to evaluate repo readiness, I moved all the scoring into a script, because I needed determinism and didn't want the scored tied to how generous the LLM was feeling that day. I've been thinking about that same division of responsibility across the agents I build. Some work needs judgment, while other work needs to be deterministic. We already know how to write software that does that. The article demonstrates this with a flaky test detector where the agent helps…