Agents Need Protocols, Not API Wrappers

AI agents are increasingly connected to software by wrapping existing APIs as tools. Wrappers are useful engineering, but they say how to call a function, not when a call is legitimate, what state must persist, how failure should be repaired, or how an interaction can be audited. We call this the interaction-contract gap . Drawing on multi-agent commitments and contemporary tool protocols, we argue for protocol interfaces: machine-readable contracts that bind capabilities to task state, context boundaries, permissions, repair semantics, and trace obligations. APIs should remain the execution substrate; the missing abstraction is an agent-facing agreement about responsible tool use.