Some failures should lead to retry: principally, LLM call failures that are not known to be permanent. However, most user code exceptions should not drive retry. We must differentiate between the two: - Provide a better default that differentiates - Allow user control over retry policy