世界经济论坛-AI代理在行动：评估和治理的基础（英）

AI Agents in Action: Foundations for Evaluation and GovernanceW H I T E P A P E RN O V E M B E R 2 0 2 5In collaboration with Capgemini Images: Adobe StockDisclaimer This document is published by the World Economic Forum as a contribution to a project, insight area or interaction. The findings, interpretations and conclusions expressed herein are a result of a collaborative process facilitated and endorsed by the World Economic Forum but whose results do not necessarily represent the views of the World Economic Forum, nor the entirety of its Members, Partners or other stakeholders.© 2025 World Economic Forum. All rights reserved. No part of this publication may be reproduced or transmitted in any form or by any means, including photocopying and recording, or by any information storage and retrieval system.ContentsForeword 4Executive summary 5Introduction 61 Evolving technical foundations of AI agents 81.1 The software architecture of an AI agent 81.2 Communication protocols and interoperability 101.3 Cybersecurity considerations 122 Foundations for AI agent evaluation and governance 132.1 Classification 142.2 Evaluation 192.3 Risk assessment 222.4 Governance considerations for AI agents: a progressive approach 253 Looking ahead: multi-agent ecosystems 29Conclusion 30Contributors 31Endnotes 34AI Agents in Action: Foundations for Evaluation and Governance2ForewordIn recent years, organizations have moved beyond predictive models and chat interfaces to experiment with artificial intelligence (AI) in more transformative ways. AI agents are now emerging as integrated collaborators in business, public services and everyday life. The adoption of AI agents could bring significant gains in efficiency, altered kinds of human-machine interaction and the advent of novel digital ecosystems.This transition faces multiple obstacles that need to be addressed. Moving from models to agents represents more than a technical milestone and requires organizations to rethink how they design, evaluate and govern advanced agentic systems. Many companies are now questioning what agents can accomplish alongside the practical steps needed to adopt and deploy them safely, responsibly and effectively.This paper was developed to help answer those questions. By mapping the evolving foundations of agentic systems, classifying their roles, identifying new ways to evaluate them and outlining progressive governance approaches, the paper offers practical guidance for leaders navigating adoption in real-world contexts.Through the AI Governance Alliance, the World Economic Forum and Capgemini are advancing this subject in collaboration with the AI community, signalling that now is the time to prepare for an agentic future. If adopters start small, iterate carefully and apply proportionate safeguards, agents can be deployed in ways that amplify human capabilities, unlock productivity and establish a foundation for more complex multi-agent ecosystems to emerge over ti