Compounding Error Rates in a Long-Running Data-Cleaning Agent

An agent is designed to clean a messy customer dataset end to end: for each of several thousand records it normalizes addresses, deduplicates against existing records, infers missing fields, and writes the result back. A single record's full cleaning trajectory involves about 8 sequential agent decisions (parse, normalize, dedupe-check, infer, validate, ...), and testing shows each individual decision is about 92% reliable in isolation.

1. Compute the approximate end-to-end reliability of one record's full trajectory, showing the arithmetic.
2. The team wants end-to-end reliability of at least 90% per record without redesigning the agent's steps. Propose the levers that would actually move this number, and roughly how much each helps.
3. Explain why "just add a retry on the whole 8-step trajectory if anything fails" is a weaker fix than improving per-step reliability, even though both can raise the end-to-end number.