An agent can issue refunds. Which safeguard suits this?
Why D is the answer
Human approval for refunds above a set limit
Refunds are high-impact actions, so they need a control outside the model, such as human approval above a limit.
In detail
- Money movement needs a limit that holds.
- Human approval above a set value does that.
- Set the limit with the business.
Use case
An illustrative scenario, not a real incident.
An agent can refund customers. Below a limit it acts alone, and above it a person approves. Mistakes stay small.
See it in different ways
sequenceDiagram
participant A as Agent
participant H as Human
participant T as Refund tool
A->>H: Refund above the limit
alt Approved
H->>T: Run
else Rejected
H-->>A: Do not run
endflowchart TD
Q{"How large is the refund?"}
Q -->|"Below the limit"| A0["Agent acts"]
Q -->|"Above the limit"| A1["Human approves"]Key point: Limits and approvals protect money.
Why the other options are wrong
- A. Raise the agent's temperature
- More randomness makes it worse.
- B. Remove the logging on that tool, so that the agent runs more quickly
- Removing logs loses your record of what happened.
- C. Add more examples to its prompt
- Examples guide the agent, and do not enforce a limit.