Checking on runs never uses up your allowance for starting them.
Headers
Responses to keyed requests carryRateLimit-Limit, RateLimit-Remaining and RateLimit-Reset (seconds until a slot frees), also sent as X-RateLimit-*, plus RateLimit-Policy.
When you hit a limit
You get HTTP429 with error.code set to rate_limited or too_many_concurrent_runs. Wait for the Retry-After header and try again. If it’s absent, back off from 0.5 seconds, doubling each time, with some random jitter.
Separately, a run can fail with internal_error when our own capacity is busy. That run is retryable, with retry_after_seconds set to 60.
Need higher limits for a production workload? Book a call and we’ll raise them.