I follow agentic workflows and tool permissions with particular attention to operating risk and maintainability. Broad permissions and weak stopping rules can turn a plausible model error into an external side effect or repeated failure.