Inference scheduler tuned for agentic micro-workloads — many short requests, no long training jobs

“Agentic workloads are highly fragmented — GPUs face many small requests instead of long jobs, breaking the long-residency assumption of training-era schedulers (https://jimmysong.io/blog/gtc-2026-ai-native-infrastructure/, https://vexxhost.com/blog/gpu-capacity-crisis-ai-infrast…”