Every team that puts an LLM agent into production writes a retry loop in the first week. It feels like reliability. It isn’t, quite.
A retry doesn’t know what finished
When a function with five external calls fails on the fifth, a retry starts again at the first. The four calls that already succeeded run again: another charge lookup, another refund, another email.
Durability records each step
A durable workflow saves the result of each step as it finishes. On a retry or a replay, finished steps return their saved results instead of running again. Only the step that failed runs again.
What it costs you
You wrap each external call in a named step. That’s the entire change, and it’s the difference between an incident and a log line.
