Adaptive quantum circuits place a classical control loop spanning over readout classification, syndrome decoding and feedforward, inside the coherence window, turning every missed deadline into a logical error rather than a delay. This talk traces that loop end-to-end and asks what it takes to orchestrate workloads around it without breaking it. Syndrome throughput scales with code distance and QPU count, but coherence fixes the latency budget. Neither Slurm nor Kubernetes was designed for a non-fungible, decohering resource whose quality drifts hourly and whose deadlines are physical . We propose a strict real-time boundary: millisecond-level orchestration handles placement and admission, then exits the loop entirely, while microsecond-and-below decoding executes on dedicated, contention-free near-QPU compute. From one QPU to a hundred, the unsettled question is whether to centralize real-time classical compute for utilization or distribute it for determinism and what breaks first when we choose wrong. We show that placement is fundamentally a prediction problem, where circuit DAG structure, non-Clifford density, live calibration state, and decoder queue depth drive resource and runtime estimates — explicitly separated from the syndrome-critical path.