Agent-Safe Design

An assistant can make valid API calls and still carry out the wrong action. Suppose a support assistant reads a ticket, drafts a reply, and sends it to the customer using a working credential. The message could still go to an address copied from an untrusted attachment, include internal notes, or reach the customer twice after a timeout.

Correct request schemas do not prevent these failures. Agent-safe design limits what an automated caller can do, binds consequential actions to explicit authority, and preserves enough execution state to recover without guessing.

This chapter develops those controls through a customer-message API.