Agent-First Tool API Paradigm for Enterprise AI Systems

A recent study published on arXiv (2605.10555) highlights five key discrepancies between traditional APIs and the needs of autonomous AI agents: reliance on exact identifiers, focus on rendering responses, assumptions of single-shot interactions, user-equivalent authorization, and unclear error semantics. To address these issues, the authors suggest an Agent-First Tool API framework that includes three components: a Six-Verb Semantic Protocol (search, resolve, preview, execute, verify, recover), a Normalized Tool Contract (NTC) that offers decision-support metadata, and a two-tier governance system that merges static capability policies with dynamic risk escalation.