What is it about?

AI agents are increasingly connected to calendars, databases, business systems, and other software through API wrappers. These wrappers tell an agent how to call a tool, but they often do not explain when the tool should be used, what permission is required, what information must be preserved, or what should happen when something goes wrong. This paper argues that agents need protocol interfaces, which are machine-readable rules that govern an interaction from beginning to end. A protocol interface would define the current task state, permitted actions, approval requirements, context boundaries, recovery steps, and the records needed to explain what happened. APIs would still perform the actual operations. Protocols would provide the missing agreement that helps an AI agent use those operations safely, consistently, and responsibly.

Featured Image

Why is it important?

Giving an AI agent access to a tool does not automatically make that tool safe for autonomous use. An API call may succeed technically while still being inappropriate. For example, an agent could schedule a meeting without approval, use outdated information, access restricted data, or complete an action without leaving a clear audit trail. Our paper identifies this problem as the interaction-contract gap. We propose protocol interfaces as a way to close this gap by making task state, permissions, context, failure recovery, and accountability explicit. This is especially important for multi-step and high-impact workflows involving people, sensitive information, or external systems. Protocol interfaces could make AI agents more reliable, easier to audit, safer to deploy, and better able to recover when plans fail.

Perspectives

This paper began with a simple concern: an AI agent may know how to call a tool without understanding the responsibilities that come with using it. Much of today’s agent ecosystem focuses on connecting models to more APIs. We believe the next challenge is more fundamental. Agents need clear agreements about what they may do, when they may do it, what approval they need, and what should happen when an action fails. We hope this paper encourages researchers, developers, and standards communities to move beyond basic tool access and design stronger interaction contracts. As agents become more autonomous, their actions must be not only executable, but also understandable, repairable, and accountable.

Tanush Sharanarthi

Read the Original

This page is a summary of: Agents Need Protocols, Not API Wrappers, ACM AI Letters, July 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3830910.
You can read the full text:

Read

Contributors

The following have contributed to this page