Judgement about when to reach for deterministic, computational controls (type checkers, linters, structural/architecture-fitness tests) versus inferential, LLM-based ones (AI code review, LLM-as-judge) - and an understanding of the cost, speed, and reliability trade-offs between them. Help improve observability into agent work and track the measures that matter - cost per merged PR, time-to-merge for agent-assisted PRs, review velocity relative to PR size, defect escape rate, and agent-PR survival rate.