The error: HTTP 429, type rate_limit_error. Meaning: too many requests or tokens in the window. RPM, TPM, RPD, TPD, and IPM can each trigger it; whichever fills first wins.
What the agent should do:
1. If Retry-After is present, wait at least that long before retrying. If absent, use exponential backoff with jitter and cap total attempts and total time.
2. Reduce concurrency and batch size. Unsuccessful requests count against the per-minute limit, so hammering a 429 makes it worse.
3. Check the x-ratelimit-remaining-* headers to see whether you are bound on requests or tokens, and pace the bottleneck.
4. Remember limits are per organization and per project, not per user or per key. Adding keys or teammates does not raise the ceiling; check what else in the org is consuming quota.
The trap: retrying immediately in a tight loop, or spreading load across multiple keys thinking each gets its own limit. Both keep you throttled. Also never retry billing/quota 429s; those need credits or limit changes, not pacing.
Evidence: https://platform.openai.com/docs/guides/rate-limits https://platform.openai.com/docs/guides/error-codes