Define what an agent may actually do

Design an authority matrix for an agent that helps with [WORKFLOW]. List read-only actions, local edits, external writes, money movement, messages to people and destructive actions separately. For each action, define trigger, scope, required evidence, confirmation rule, retry safety and audit record. Give three realistic ambiguous examples and show whether the agent should act, pause or ask. Produce an operator-readable policy plus a machine-friendly table. Do not invent permissions or treat a request to 'handle it' as unlimited authority. State which tasks remain human-owned and how the agent reports uncertainty.