Comparing.
Your agent wrote both drafts, and a model asked to judge its own work tends to like it. A person reads the two side by side and picks the one that should go out.
- An outside readerSomeone who did not write either candidate.
- One of the twoThe answer is a choice between what you sent, not a rewrite.
- Priced by the mistakeA reply to an angry customer is worth more certainty than a tagline.
- Says when it is a tieIf people split, you are told the two are as good as each other, and refunded.
Which reply should go to a customer charged twice?
What happened
Two readers who wrote neither chose the same reply, and that is the one sent.
How it works
Step by step.
The question is the one your agent could not answer about its own work: which of these is better.
Real-world example
The reply that keeps the customer.A support agent drafts two answers to a customer who was charged twice. Scored by the model that wrote them, the stiff one edges ahead. Read by people, the apology wins every time, and that is the one that goes out.
For crypto agents
Where this saves a transaction.
An on-chain payment cannot be taken back, so the moment before the agent signs is the moment to ask.
deadline_ms to the window you actually have: if nobody answers in time you are refunded, and the agent takes its safe path, which is not to sign.