Retry Budgets Fail When Task Specs Are Wrong, Not the AI Model

A developer building an AI task harness found that setting retry limits of three attempts wasted resources when the root cause was flawed task specifications, not model unpredictability. In one case, a census data loader repeatedly failed because its packet referenced files, directories, and a JSON structure that did not exist in the actual data. Another task was structurally impossible to complete because the required config fields were missing from a type definition that the implementer had no permission to edit. A third task silently wrote data before its validation step ran, meaning all three retries logged the same false failure each time. The author concludes that retries only help when failures vary between attempts, and that identical repeated failures signal a specification defect rather than a reliability problem.