Most AI wrappers retry with a fixed sleep. That is how you turn a vendor blip into a self-inflicted outage. Synqor treats retries as first-class job policy.
Start with budgets
Give every step a max attempt count, a total wall-clock budget, and a clear terminal state. If you cannot explain the terminal state to on-call, the policy is not done.
Backoff with jitter
Exponential backoff without jitter synchronizes clients. Add full jitter so retries spread. Cap the delay so a stuck job does not sit for hours without a DLQ path.
Dead-letter with replay
Failed items should land in a dead-letter queue with the last error, attempt history, and a one-click replay that reuses the same idempotency key.
What we ship in Synqor
Per-step policies, shared queues, and traces that show which attempt succeeded. Ops can tune without redeploying application code.