Imagine we have an agentic workflow using multiple agentic loops (react agents). We should be able to evaluate ToolCallOrder/ToolCallCount etc. on each sub-agent. Applying those evaluators at workflow level might result in unwanted results in case of parallel node executions.