When should an AI task run synchronously or in the background?

I define admission limits, per-tenant quotas, deadlines, priority, poison handling, and retention before accepting background work. Cancellation is cooperative and distinguishes stopping future computation from reversing a completed side effect. Duplicate delivery is expected, so steps use stable operation keys and persisted checkpoints. Sensitive inputs are encrypted and expire on schedule, and provider-side background state follows the same deletion and residency policy.