Human-In-The-Loop Approval Workflow Specifier

ROLE: You are a workflow designer balancing agent autonomy with human oversight.

CONTEXT: The agent automates [PROCESS]. Some actions are reversible and cheap; others are irreversible or high-stakes, specifically [HIGH_STAKES_ACTIONS]. Stakeholders are [APPROVERS].

TASK: Define the human-in-the-loop policy.
1. Categorize every action the agent can take as auto-execute, notify-then-execute, or require-approval. Justify the cutoffs.
2. For each require-approval action, specify exactly what is shown to the approver: the proposed action, the reasoning, the predicted effect, and the cost of being wrong.
3. Define the approval interface: options the human can pick (Approve / Reject / Modify) and what each does.
4. Specify timeout behavior if no human responds within [SLA].
5. Define how the agent records the decision and resumes the workflow.

OUTPUT FORMAT: (a) Action-tier table; (b) an approval-request card template with all required fields; (c) resume logic per decision type.

CONSTRAINTS: Default to caution: when in doubt, require approval. The approval card must let a busy human decide in under 30 seconds. Never auto-execute an irreversible action.