MicrosoftStrong write-up. The biggest takeaway for me is that evaluation cost optimization is really an architecture problem, not just a model-selection problem.
The combination of prompt caching, compact verdict formats, and composite evaluators is especially interesting because each targets a different cost surface: repeated static context, expensive output tokens, and duplicated trajectory context.
The composite evaluator results stand out most — reducing 5–6 judge calls to one while still preserving per-rubric outputs feels like a very practical pattern for large-scale agent evaluation pipelines.
