Proposes treating entire agent workflows — not individual inference calls — as the first-class schedulable unit on GPU clusters. Current request-level schedulers discard gigabytes of intermediate state between steps, inflating end-to-end latency 3-8x for compound AI workloads where agents execute tens to hundreds of chained LLM calls per task. SAGA shifts to program-level scheduling, preserving KV cache and intermediate state across the workflow lifetime.