Retrying whole ten-step agent jobs wasted 4,143 model calls ($42.26) on 2,000 jobs, lost 75 jobs, and sent a duplicate email for 273 jobs (our lab, 2026).
Saving a checkpoint after each step cut wasted calls to 822 ($8.38) and lost no jobs, with each write taking 6.8 microseconds on SQLite, but 21 jobs still sent a duplicate email.
An idempotency key on each side effect, such as job ID plus step number, removed the remaining duplicates, and a workflow engine is only needed for long waits, approvals or parallel branches.