Cap the cost of an agent's worst day

Give an agent a tool that can spend money or delete data, and the question isn't whether it will make a mistake. It's what the mistake costs.

Guardrails are how you cap that cost. A few that earn their keep:

1. 1Least privilege. Give each tool the smallest permission it needs. Read-only unless writing is the point.
2. 2Confirmation on the irreversible. Anything you can't undo, like payments, deletes and outbound messages, waits for a human yes.
3. 3Budgets. A hard limit on steps, tokens and money per run, so a loop can't run all night.
4. 4Dry runs. Let the agent show what it would do before it does it.
5. 5A log of every action, with the inputs. When something goes wrong you need the trail.

The trick with human approval is to ask rarely and ask well. If the agent asks permission for every small thing, people click yes without reading, and you've built a rubber stamp. Save it for the few actions that matter, and show exactly what will happen.

A good test: if this agent had a bad day, what's the worst thing it could do with the access it has right now?

What's the one action your agent can take that you'd want a human to approve first?