AI代理在行动：评估和治理的基础（英）

AI Agents in Action:Foundations for Evaluationand Governance W H I T EP A P E RN O V E M B E R2 0 2 5 Contents Foreword4 Executive summary5 Introduction6 1Evolving technical foundations of AI agents8 1.1The software architecture of an AI agent81.2Communication protocols and interoperability101.3Cybersecurity considerations12 2Foundations for AI agent evaluation and governance13 2.1Classification142.2Evaluation192.3Risk assessment222.4Governance considerations for AI agents: a progressive approach25 3Looking ahead: multi-agent ecosystems29 Conclusion30 Contributors31 Endnotes34 Disclaimer This document is published by theWorld Economic Forum as a contributionto a project, insight area or interaction.The findings, interpretations andconclusions expressed herein are a resultof a collaborative process facilitated andendorsed by the World Economic Forumbut whose results do not necessarilyrepresent the views of the World EconomicForum, nor the entirety of its Members,Partners or other stakeholders. ©2025 World Economic Forum. All rightsreserved. No part of this publication maybe reproduced or transmitted in any formor by any means, including photocopyingand recording, or by any informationstorage and retrieval system. Foreword Cathy LiHead, Centre for AIExcellence, Member ofthe Executive Committee,World Economic Forum Roshan GyaChief Executive Officer,Capgemini Invent Through the AI Governance Alliance, the WorldEconomic Forum and Capgemini are advancingthis subject in collaboration with the AI community,signalling that now is the time to prepare for anagentic future. If adopters start small, iteratecarefully and apply proportionate safeguards,agents can be deployed in ways that amplifyhuman capabilities, unlock productivity andestablish a foundation for more complex multi-agentecosystems to emerge over time. Unless a carefuland deliberate approach to adoption is adopted,untested use cases could outpace oversight andlead to misaligned incentives, emergent risks andloss of public trust. In recent years, organizations have moved beyondpredictive models and chat interfaces to experimentwith artificial intelligence (AI) in more transformativeways. AI agents are now emerging as integratedcollaborators in business, public services andeveryday life. The adoption of AI agents couldbring significant gains in efficiency, altered kinds ofhuman-machine interaction and the advent of noveldigital ecosystems. This transition faces multiple obstacles that need to beaddressed. Moving from models to agents representsmore than a technical milestone and requiresorganizations to rethink how they design, evaluate andgovern advanced agentic systems. Many companiesare now questioning what agents can accomplishalongside the practical steps needed to adopt anddeploy them safely, responsibly and effectively. As with any transformative technology, theopportunities presented by AI agents must beaccompanied by a responsibility to guide theirdevelopment and deployment with care. Throughcross-functional efforts and collaborativegovernance, AI agents can be integrated in waysthat amplify human ingenuity, promote innovationand improve overall quality of life. This paper is astep in that direction, offering guidance to help earlyadopters navigate the complex and often unevenpath of AI agent adoption. This paper was developed to help answer thosequestions. By mapping the evolving foundationsof agentic systems, classifying their roles,identifying new ways to evaluate them and outliningprogressive governance approaches, the paperoffers practical guidance for leaders navigatingadoption in real-world contexts. Executive summary This paper explores the emergence of AIagents, outlining their technical foundations,classification, evaluation and governance tosupport safe and effective adoption. classification that differentiates agents by theirrole,autonomy, authority, predictability and operationalcontext. Thirdly, it suggests a progressive governanceapproach that directly connects evaluation andsafeguards to an agent’s task scope anddeployment environment. This report has been tailored mainly for adoptersof AI agents, including decision-makers, technicalleaders and practitioners seeking to integrate AIagents into organizational workflows and services. While AI agents are gaining traction, there remainslimited guidance on how to design, test andoversee them responsibly. This paper aims to helpfill that gap by providing a structured foundation forthe safe and effective deployment of these systems. Together, these elements guide adopters withaconceptual blueprint for moving from experimentationto deployment. The report highlights the importanceof aligning adoption with evaluation and governancepractices to ensure that AI agents are successfullydeployed while trust, safety and accountabilityare maintained. The paper makes three key contributions. Firstly,it covers the technical foundations of AI agents,including their architect