Decide what an agent should escalate to a person and what it should just do

You are an automation reviewer who has cleaned up after agents that asked permission for everything and agents that asked for nothing.

What the agent does, step by step: {{paste the workflow}}. The actions it can take on real systems: {{list tools, APIs, accounts, spend limits}}. What has already gone wrong or nearly gone wrong: {{paste, even anecdotes}}.

Return five sections:
1. ACTION LEDGER - table: Action | Reversible? | Worst realistic outcome | Auto / Notify-after / Approve-first. Every action from my list appears exactly once.
2. THE APPROVE-FIRST SET - keep it to the smallest set that covers the irreversible and expensive actions; justify each in one line.
3. THE HANDOFF - for each approve-first action, what the agent must show the human: the exact fields, and the default if nobody answers in an hour.
4. NOISE CHECK - which of my current notifications a person will start ignoring within a week, and what to delete.
5. THE TRIPWIRE - one condition that should stop the whole run, not just one action.

Rules: use only the tools I listed, do not invent approval UI I do not have, no talk of 'responsible AI' in the abstract.