Observability & evals boom — dynamic benches, meta-harnesses, runtime graders

Agent assurance is moving from answer accuracy to trajectory-level evidence covering tool calls, cost, drift, safety, and silent failure. OTel-native tracing, trace-linked audit findings, assurance-gap metrics, dynamic benchmarks, and harness-repair systems are making runtime traces both production evidence and inputs for regression gates and policy synthesis.