Role in the stack
Single-call tool use is insufficient for many tasks. The runtime, not the model, should own retries, backoff, idempotency, timeout budgets, batching, circuit breaking, webhook correlation and compensation. Where workflows matter, describe them explicitly with Arazzo rather than leaving the model to infer every step.
Evidence
ToolSandbox and τ-bench show state dependence and consistency remain weak points — which is exactly why orchestration belongs in infrastructure.
Relates to: Workflow engine, OpenAI Agents SDK.
Terms
Glossary terms used here: Compensation, Idempotency, Tool use, Workflow engine.