Many background jobs call an external system on behalf of separate accounts, tenants, or installations. The external system allows parallel calls across different accounts, but it does not allow two calls for the same account to run at the same time. That is not a global rate limit. It is concurrency by key: the key might be account_id , tenant_id , or another argument that identifies the shared quota or state boundary. You want the worker pool busy across many accounts, while each account stays serial. Without that guard, two workers eventually pick up work for the same account in parallel. The external system may throttle the account, reject the second call, or leave you with a partial update to reconcile. The usual fixes are external locks, one queue per account, or retry/backoff logic around every call. They can work, but they add another coordination layer to the job system. Pynenc's orchestrator already tracks running invocations and their arguments.…