dualdust

The timeout example is the part I'd make observable at the business-action level, not just the agent-run level. Give each intended action an idempotency key and keep requested, committed, confirmed, and compensated as separate states. A retry is safe only when you know which state you're in. I'd also keep the model out of retry decisions for deterministic API failures; a small state machine should own those. What context do you send when the run finally escalates to a human?

I like the focus on traffic shape. I would add one more comparison layer: cost per accepted outcome, not only cost per model call. For RAG or agent workflows, a cheaper model can still need more context, retries, reranking, or eval passes. That can make raw dollars-per-million-tokens misleading. The budget I would want to see is workflow-level: input, output, retrieval/context assembly, fallback/retry, and eval/judge calls.