A queue is a mechanism for controlling demand and recovering work after interruption; an agent orchestrator is a mechanism for deciding and coordinating what to do. Confusing the two leaves teams with expensive language-model loops but no dependable record of business progress.
How it works
Temporal documentation describes workflows, workers and activities, while event platforms document sources that trigger downstream jobs. Use a queue to absorb bursts, apply backpressure and retry temporary failures. Use durable workflow state when operations span several services, approval periods and compensation steps. Queue delivery itself is not proof that a target API executed exactly once, so every external write needs its own idempotency strategy.
A practical, bounded workflow
For incoming invoices, a listener accepts an upload and enqueues only a verified pointer and event ID. A worker fetches and validates the file, then runs an extraction model with a constrained budget. A workflow checkpoint stores the proposed line items. A human approves them before a separate accounting worker performs a write. If the worker crashes after the write, its successor queries the accounting system by business key instead of creating another invoice.
Failure modes and limits
Poorly configured retries turn a temporary outage into a queue storm. Apply bounded exponential backoff, a dead-letter lane for persistent errors, and per-tenant concurrency limits. Distinguish failures caused by invalid input from network failures: the former should not consume unlimited retries. Cancellation should stop future work but cannot retroactively revoke confirmed side effects.
Deployment controls
Assign responsibility for each transition: scheduled, running, awaiting evidence, awaiting approval, writing, reconciled, canceled or unresolved. The controller, not a free-form language response, decides whether a next action is authorized. Persist operation IDs and deadlines so a restart does not erase obligations or repeat committed effects.
When to choose this approach
An ordinary background job runner can be sufficient for one idempotent function. A durable orchestrator earns its overhead when workflows depend on human decisions, third-party state and multi-day waits. Avoid equating ‘supports retries’ with end-to-end correctness; each connector’s semantics must be measured.
What to evaluate before production
Simulate a duplicate event, worker termination, service outage, expired approval and partial remote commit. Record queue age, retry distribution, dead-letter count and reconciled business outcomes. Verify the same input yields one approved external effect and clear unresolved alerts when exact completion cannot be determined.
Version, market and tier caution (25 October 2026): This is a technical and operational guide, not a live entitlement, pricing or geographic-availability certification. Numeric commercial limits and undocumented release dates are intentionally omitted. Reconfirm the exact account, deployment region, access policy and vendor release before procurement or publication.