Workflow action nodes that dispatch to external backends like GitHub or webhooks fail fast on transient errors (429, 502, 503, 504) or provider outages, with no retry or backoff. The run is marked failed, so someone has to notice it and manually re-trigger once the provider recovers. Configurable per-node retries (attempt count, exponential backoff) and optional queueing of dispatches during downstream degradation would let runs recover on their own, with opt-in controls so non-idempotent actions aren't retried unsafely. Without this, a brief third-party incident breaks developer self-service and pushes manual monitoring and re-runs onto end users.