Dynamic Workflows in Azure Functions Hosted Skills: the LLM plans a DAG once, Durable Functions run it — 56–93% fewer tokens

Public preview, Python-only. Microsoft's benchmark on gpt-5.4-mini shows token use down 56–93% and latency down 77–95% versus looping the model per step. Limits: 50 nodes, 10 parallel tasks, 24-hour waits.