One edge case to specify: if a 401 can be fixed by refreshing credentials, allow one refresh and replay of the original request, then fail rather than entering the generic retry loop again. For non-idempotent writes, require an idempotency key before retrying a timeout or 5xx; keep the attempt cap inside a total deadline too. I work with BOMTOBE, the team behind Orke, so I’m disclosing that affiliation while sharing this agent-workflow feedback.