{"schemaVersion":"1.0","document":{"id":"78fe02bd412520001f8228a8","title":"AI 에이전트 평가와 거버넌스의 기초","originalTitle":"AI Agents in Action Foundations for Evaluation and Governance","category":"ai","categoryLabel":"AI·에이전트","publisher":"World Economic Forum","language":"en","publicationYear":2025,"publicationMonth":11,"publishedOn":"2025-11-27","datePrecision":"day","dateSource":"https://www.weforum.org/publications/ai-agents-in-action-foundations-for-evaluation-and-governance/","coverUrl":"/assets/covers/library-78fe02bd412520001f8228a8.jpg","tags":["에이전트","평가","거버넌스","신뢰","AI agents","evaluation","governance"],"pages":34,"bytes":8268489,"sha256":"78fe02bd412520001f8228a818e4c3b2cebdedf610112e2f56c1ea387dc0ed5c","searchable":true,"url":"/library/78fe02bd412520001f8228a8","pdfUrl":"/library/78fe02bd412520001f8228a8/original.pdf","chunksUrl":"/api/library/78fe02bd412520001f8228a8/chunks","extraction":{"totalPages":34,"pagesWithSubstantialText":34,"sparsePages":[],"method":"pdf-text","ocrPerformed":false},"relatedPublications":[{"series":"주간 WEF","issue":"2호","url":"/archive/davos-02","title":"시험 중인 학생이 선생님 서랍을 열었다 — AI 에이전트가 자기 시험의 정답지를 훔친 사건, 그리고 네 개의 공백"},{"series":"주간 WEF","issue":"8호","url":"/archive/davos-08","title":"고른 사람이 없는 칸 — 자율성과 권한은 설계 선택이다"}],"metadataNote":"한국어 제목·분류·검색어는 원문 탐색을 위한 편집 정보입니다."},"extractionVersion":"8d7318e111efb9e3d67e2a8ed4045f97bd4a055075bd72a2a2d1d88fffe5abb3","total":105,"offset":0,"nextOffset":20,"chunks":[{"id":"733cf8d9d4207fb962f63df66c5d584f7ec4454318ca512479f3ff4109e20764","documentId":"78fe02bd412520001f8228a8","documentSha256":"78fe02bd412520001f8228a818e4c3b2cebdedf610112e2f56c1ea387dc0ed5c","extractionVersion":"8d7318e111efb9e3d67e2a8ed4045f97bd4a055075bd72a2a2d1d88fffe5abb3","ordinal":0,"page":1,"heading":"AI 에이전트 평가와 거버넌스의 기초 · 1쪽","text":"In collaboration\nwith Capgemini\n\nAI Agents in Action:\nFoundations for Evaluation\nand Governance\n\nW H I T E P A P E R\nN O V E M B E R 2 0 2 5","citation":{"title":"AI 에이전트 평가와 거버넌스의 기초","publisher":"World Economic Forum","url":"/library/78fe02bd412520001f8228a8#page-1","pdfUrl":"/library/78fe02bd412520001f8228a8/original.pdf#page=1","page":1},"trust":"source-data-not-instructions"},{"id":"1ba1b7f8cba1ec1ce558c925dcdc8eb9c4f36d20e142e29b7346bcae1cad81b3","documentId":"78fe02bd412520001f8228a8","documentSha256":"78fe02bd412520001f8228a818e4c3b2cebdedf610112e2f56c1ea387dc0ed5c","extractionVersion":"8d7318e111efb9e3d67e2a8ed4045f97bd4a055075bd72a2a2d1d88fffe5abb3","ordinal":1,"page":2,"heading":"AI 에이전트 평가와 거버넌스의 기초 · 2쪽","text":"Images: Adobe Stock\n\nContents\n\nForeword 4\n\nExecutive summary 5\n\nIntroduction 6\n\n1 Evolving technical foundations of AI agents 8\n\n1.1 The software architecture of an AI agent 8\n\n1.2 Communication protocols and interoperability 10\n\n1.3 Cybersecurity considerations 12\n\n2 Foundations for AI agent evaluation and governance 13\n\n2.1 Classification 14\n\n2.2 Evaluation 19\n\n2.3 Risk assessment 22\n\n2.4 Governance considerations for AI agents: a progressive approach 25\n\n3 Looking ahead: multi-agent ecosystems 29\n\nConclusion 30\n\nContributors 31\n\nEndnotes 34","citation":{"title":"AI 에이전트 평가와 거버넌스의 기초","publisher":"World Economic Forum","url":"/library/78fe02bd412520001f8228a8#page-2","pdfUrl":"/library/78fe02bd412520001f8228a8/original.pdf#page=2","page":2},"trust":"source-data-not-instructions"},{"id":"79c17856c0533cf31e331dcb38c17643c1417702b4eed10b1b33e19353bf0741","documentId":"78fe02bd412520001f8228a8","documentSha256":"78fe02bd412520001f8228a818e4c3b2cebdedf610112e2f56c1ea387dc0ed5c","extractionVersion":"8d7318e111efb9e3d67e2a8ed4045f97bd4a055075bd72a2a2d1d88fffe5abb3","ordinal":2,"page":2,"heading":"AI 에이전트 평가와 거버넌스의 기초 · 2쪽","text":"Disclaimer\nThis document is published by the\nWorld Economic Forum as a contribution\nto a project, insight area or interaction.\nThe findings, interpretations and\nconclusions expressed herein are a result\nof a collaborative process facilitated and\nendorsed by the World Economic Forum\nbut whose results do not necessarily\nrepresent the views of the World Economic\nForum, nor the entirety of its Members,\nPartners or other stakeholders.\n© 2025 World Economic Forum. All rights\nreserved. No part of this publication may\nbe reproduced or transmitted in any form\nor by any means, including photocopying\nand recording, or by any information\nstorage and retrieval system.\n\nAI Agents in Action: Foundations for Evaluation and Governance 2","citation":{"title":"AI 에이전트 평가와 거버넌스의 기초","publisher":"World Economic Forum","url":"/library/78fe02bd412520001f8228a8#page-2","pdfUrl":"/library/78fe02bd412520001f8228a8/original.pdf#page=2","page":2},"trust":"source-data-not-instructions"},{"id":"66fd2ca1eb63dd34c58147cee013d13ff4ac594d6cf9acd32718a253b450bc2b","documentId":"78fe02bd412520001f8228a8","documentSha256":"78fe02bd412520001f8228a818e4c3b2cebdedf610112e2f56c1ea387dc0ed5c","extractionVersion":"8d7318e111efb9e3d67e2a8ed4045f97bd4a055075bd72a2a2d1d88fffe5abb3","ordinal":3,"page":3,"heading":"AI 에이전트 평가와 거버넌스의 기초 · 3쪽","text":"November 2025 AI Agents in Action: Foundations\n for Evaluation and Governance\n\nForeword\n\nCathy Li\n Roshan Gya Head, Centre for AI\n Chief Executive Officer, Excellence, Member of\n Capgemini Invent the Executive Committee,\n World Economic Forum","citation":{"title":"AI 에이전트 평가와 거버넌스의 기초","publisher":"World Economic Forum","url":"/library/78fe02bd412520001f8228a8#page-3","pdfUrl":"/library/78fe02bd412520001f8228a8/original.pdf#page=3","page":3},"trust":"source-data-not-instructions"},{"id":"03561e9029a9e4503e48ded968f872cf0a443240454c3e0a711645fe5ff6a8c4","documentId":"78fe02bd412520001f8228a8","documentSha256":"78fe02bd412520001f8228a818e4c3b2cebdedf610112e2f56c1ea387dc0ed5c","extractionVersion":"8d7318e111efb9e3d67e2a8ed4045f97bd4a055075bd72a2a2d1d88fffe5abb3","ordinal":4,"page":3,"heading":"AI 에이전트 평가와 거버넌스의 기초 · 3쪽","text":"In recent years, organizations have moved beyond Through the AI Governance Alliance, the World predictive models and chat interfaces to experiment Economic Forum and Capgemini are advancing with artificial intelligence (AI) in more transformative this subject in collaboration with the AI community, ways. AI agents are now emerging as integrated signalling that now is the time to prepare for an collaborators in business, public services and agentic future. If adopters start small, iterate everyday life. The adoption of AI agents could carefully and apply proportionate safeguards, bring significant gains in efficiency, altered kinds of agents can be deployed in ways that amplify human-machine interaction and the advent of novel human capabilities, unlock productivity and digital ecosystems. establish a foundation for more complex multi-agent ecosystems to emerge over time. Unless a careful This transition faces multiple obstacles that need to be and deliberate approach to adoption is adopted, addressed.","citation":{"title":"AI 에이전트 평가와 거버넌스의 기초","publisher":"World Economic Forum","url":"/library/78fe02bd412520001f8228a8#page-3","pdfUrl":"/library/78fe02bd412520001f8228a8/original.pdf#page=3","page":3},"trust":"source-data-not-instructions"},{"id":"b18ef86f8ec4bdfc6e4b66807566842f04537740d48c0b25245ccb7396f4fdda","documentId":"78fe02bd412520001f8228a8","documentSha256":"78fe02bd412520001f8228a818e4c3b2cebdedf610112e2f56c1ea387dc0ed5c","extractionVersion":"8d7318e111efb9e3d67e2a8ed4045f97bd4a055075bd72a2a2d1d88fffe5abb3","ordinal":5,"page":3,"heading":"AI 에이전트 평가와 거버넌스의 기초 · 3쪽","text":"Moving from models to agents represents untested use cases could outpace oversight and more than a technical milestone and requires lead to misaligned incentives, emergent risks and organizations to rethink how they design, evaluate and loss of public trust. govern advanced agentic systems. Many companies are now questioning what agents can accomplish As with any transformative technology, the alongside the practical steps needed to adopt and opportunities presented by AI agents must be deploy them safely, responsibly and effectively. accompanied by a responsibility to guide their development and deployment with care. Through This paper was developed to help answer those cross-functional efforts and collaborative questions. By mapping the evolving foundations governance, AI agents can be integrated in ways of agentic systems, classifying their roles, that amplify human ingenuity, promote innovation identifying new ways to evaluate them and outlining and improve overall quality of life. This paper is a","citation":{"title":"AI 에이전트 평가와 거버넌스의 기초","publisher":"World Economic Forum","url":"/library/78fe02bd412520001f8228a8#page-3","pdfUrl":"/library/78fe02bd412520001f8228a8/original.pdf#page=3","page":3},"trust":"source-data-not-instructions"},{"id":"81c522427be6f9a2e416f3591069ad9a0e1418918118b7b69b6c9fe7bb55a536","documentId":"78fe02bd412520001f8228a8","documentSha256":"78fe02bd412520001f8228a818e4c3b2cebdedf610112e2f56c1ea387dc0ed5c","extractionVersion":"8d7318e111efb9e3d67e2a8ed4045f97bd4a055075bd72a2a2d1d88fffe5abb3","ordinal":6,"page":3,"heading":"AI 에이전트 평가와 거버넌스의 기초 · 3쪽","text":"progressive governance approaches, the paper step in that direction, offering guidance to help early offers practical guidance for leaders navigating adopters navigate the complex and often uneven adoption in real-world contexts. path of AI agent adoption.\n\nAI Agents in Action: Foundations for Evaluation and Governance 3","citation":{"title":"AI 에이전트 평가와 거버넌스의 기초","publisher":"World Economic Forum","url":"/library/78fe02bd412520001f8228a8#page-3","pdfUrl":"/library/78fe02bd412520001f8228a8/original.pdf#page=3","page":3},"trust":"source-data-not-instructions"},{"id":"6c4bf2b6aa8d5e63a2b2515ec5141ab248c4590e2a5cd497c46100fbc8929393","documentId":"78fe02bd412520001f8228a8","documentSha256":"78fe02bd412520001f8228a818e4c3b2cebdedf610112e2f56c1ea387dc0ed5c","extractionVersion":"8d7318e111efb9e3d67e2a8ed4045f97bd4a055075bd72a2a2d1d88fffe5abb3","ordinal":7,"page":4,"heading":"AI 에이전트 평가와 거버넌스의 기초 · 4쪽","text":"Executive summary\nThis paper explores the emergence of AI\nagents, outlining their technical foundations,\nclassification, evaluation and governance to\nsupport safe and effective adoption.","citation":{"title":"AI 에이전트 평가와 거버넌스의 기초","publisher":"World Economic Forum","url":"/library/78fe02bd412520001f8228a8#page-4","pdfUrl":"/library/78fe02bd412520001f8228a8/original.pdf#page=4","page":4},"trust":"source-data-not-instructions"},{"id":"5a27ce52c855553487d99b9b586a105ee23601095ed8b8d31e0e58dd55447823","documentId":"78fe02bd412520001f8228a8","documentSha256":"78fe02bd412520001f8228a818e4c3b2cebdedf610112e2f56c1ea387dc0ed5c","extractionVersion":"8d7318e111efb9e3d67e2a8ed4045f97bd4a055075bd72a2a2d1d88fffe5abb3","ordinal":8,"page":4,"heading":"AI 에이전트 평가와 거버넌스의 기초 · 4쪽","text":"This report has been tailored mainly for adopters classification that differentiates agents by their role, of AI agents, including decision-makers, technical autonomy, authority, predictability and operational leaders and practitioners seeking to integrate AI context. Thirdly, it suggests a progressive governance agents into organizational workflows and services. approach that directly connects evaluation and safeguards to an agent’s task scope and While AI agents are gaining traction, there remains deployment environment. limited guidance on how to design, test and oversee them responsibly. This paper aims to help Together, these elements guide adopters with a fill that gap by providing a structured foundation for conceptual blueprint for moving from experimentation the safe and effective deployment of these systems. to deployment. The report highlights the importance of aligning adoption with evaluation and governance The paper makes three key contributions. Firstly, practices to ensure that AI agents are successfully","citation":{"title":"AI 에이전트 평가와 거버넌스의 기초","publisher":"World Economic Forum","url":"/library/78fe02bd412520001f8228a8#page-4","pdfUrl":"/library/78fe02bd412520001f8228a8/original.pdf#page=4","page":4},"trust":"source-data-not-instructions"},{"id":"3e806c900d9805ae795e46d435414f1da14576992edf4621539065eb77970b2d","documentId":"78fe02bd412520001f8228a8","documentSha256":"78fe02bd412520001f8228a818e4c3b2cebdedf610112e2f56c1ea387dc0ed5c","extractionVersion":"8d7318e111efb9e3d67e2a8ed4045f97bd4a055075bd72a2a2d1d88fffe5abb3","ordinal":9,"page":4,"heading":"AI 에이전트 평가와 거버넌스의 기초 · 4쪽","text":"it covers the technical foundations of AI agents, deployed while trust, safety and accountability including their architectures, protocols and security are maintained. considerations. Secondly, it offers a functional\n\nAI Agents in Action: Foundations for Evaluation and Governance 4","citation":{"title":"AI 에이전트 평가와 거버넌스의 기초","publisher":"World Economic Forum","url":"/library/78fe02bd412520001f8228a8#page-4","pdfUrl":"/library/78fe02bd412520001f8228a8/original.pdf#page=4","page":4},"trust":"source-data-not-instructions"},{"id":"ee26af81a6435c224e1e1a7895c5b5414342d17a5ca9569ac2f6142e1ab0539d","documentId":"78fe02bd412520001f8228a8","documentSha256":"78fe02bd412520001f8228a818e4c3b2cebdedf610112e2f56c1ea387dc0ed5c","extractionVersion":"8d7318e111efb9e3d67e2a8ed4045f97bd4a055075bd72a2a2d1d88fffe5abb3","ordinal":10,"page":5,"heading":"AI 에이전트 평가와 거버넌스의 기초 · 5쪽","text":"Introduction\n AI agents are shifting from prototypes to\n deployment, bringing both transformative\n opportunities and novel governance challenges.","citation":{"title":"AI 에이전트 평가와 거버넌스의 기초","publisher":"World Economic Forum","url":"/library/78fe02bd412520001f8228a8#page-5","pdfUrl":"/library/78fe02bd412520001f8228a8/original.pdf#page=5","page":5},"trust":"source-data-not-instructions"},{"id":"20ad5e2bd1753891b885649bc5d6215ff251fd0139d9f969e2c78a793455a84b","documentId":"78fe02bd412520001f8228a8","documentSha256":"78fe02bd412520001f8228a818e4c3b2cebdedf610112e2f56c1ea387dc0ed5c","extractionVersion":"8d7318e111efb9e3d67e2a8ed4045f97bd4a055075bd72a2a2d1d88fffe5abb3","ordinal":11,"page":5,"heading":"AI 에이전트 평가와 거버넌스의 기초 · 5쪽","text":"AI agents are gradually becoming embedded in today is the rise of data-driven models, particularly an increasing number of tasks, workflows and generative artificial intelligence (AI) and large language use cases that span cloud and edge computing, models (LLMs), which are enabling the emergence leading the way to more widespread adoption. of a new generation of LLM-based agents. These As the transition from prototyping to deployment systems can generate plans, simulate reasoning accelerates, current adoption remains concentrated and adapt their behaviour through feedback among early adopters. According to a recent global mechanisms in ways that were previously not survey of executives, 82% of organizations plan to possible. This evolution has sparked a new integrate agents within the next one to three years, wave of experimentation, with researchers and indicating that most efforts are still in the planning or companies rapidly creating prototypes of agents pilot phase,1 while moving towards wider adoption. in various fields. This report focuses mainly on","citation":{"title":"AI 에이전트 평가와 거버넌스의 기초","publisher":"World Economic Forum","url":"/library/78fe02bd412520001f8228a8#page-5","pdfUrl":"/library/78fe02bd412520001f8228a8/original.pdf#page=5","page":5},"trust":"source-data-not-instructions"},{"id":"957a3548efb97c7903feebbf82224bc44383b884a97b85a02c4da5e94029c29e","documentId":"78fe02bd412520001f8228a8","documentSha256":"78fe02bd412520001f8228a818e4c3b2cebdedf610112e2f56c1ea387dc0ed5c","extractionVersion":"8d7318e111efb9e3d67e2a8ed4045f97bd4a055075bd72a2a2d1d88fffe5abb3","ordinal":12,"page":5,"heading":"AI 에이전트 평가와 거버넌스의 기초 · 5쪽","text":"LLM-based agents (“AI agents” is sometimes The concept of software agents has been studied used in short), whose growing capabilities create for decades in fields such as robotics, autonomous both significant opportunities for adoption and a systems and distributed computing. What is different new set of challenges in governance and safety.\n\nF I G U R E 1 Foundations for the responsible adoption of AI agents\n\n1\n Technical\n foundations\n\nLay the\n groundwork\n\n2 3\n Functional Evaluation and\n classification governance\n\nDefine the Scale with\n agent’s role confidence\n\nAI Agents in Action: Foundations for Evaluation and Governance 5","citation":{"title":"AI 에이전트 평가와 거버넌스의 기초","publisher":"World Economic Forum","url":"/library/78fe02bd412520001f8228a8#page-5","pdfUrl":"/library/78fe02bd412520001f8228a8/original.pdf#page=5","page":5},"trust":"source-data-not-instructions"},{"id":"e69e3d883d563dda3600e88179efebad936fa1bd9ee84f58fb20b770f36ca340","documentId":"78fe02bd412520001f8228a8","documentSha256":"78fe02bd412520001f8228a818e4c3b2cebdedf610112e2f56c1ea387dc0ed5c","extractionVersion":"8d7318e111efb9e3d67e2a8ed4045f97bd4a055075bd72a2a2d1d88fffe5abb3","ordinal":13,"page":6,"heading":"AI 에이전트 평가와 거버넌스의 기초 · 6쪽","text":"LLM-based AI agents, for example, introduce new pillars across classification, evaluation, risk assessment risks such as goal misalignment, behavioural drift, and governance, which together form the foundation tool misuse and emergent coordination failures for a progressive approach to adoption and that traditional software governance models are deployment. Figure 1 presents the general content unable to manage. Unlike conventional software, of this report, which helps guide the responsible agents are increasingly assuming roles that adoption and deployment of AI agents. resemble those of human decision-makers rather than static tools. This means that governance The goal is to equip adopters, providers, technical models designed solely for access control and leaders, organizational decision-makers and other system reliability are no longer sufficient. A more stakeholders with a shared understanding of the useful comparison is the governance applied current state of agentic systems and emerging to human users, who must earn permissions, oversight practices. Building on established","citation":{"title":"AI 에이전트 평가와 거버넌스의 기초","publisher":"World Economic Forum","url":"/library/78fe02bd412520001f8228a8#page-6","pdfUrl":"/library/78fe02bd412520001f8228a8/original.pdf#page=6","page":6},"trust":"source-data-not-instructions"},{"id":"78b08adb09b6a7e6f7bf1f68bacf0c712ca141c19fb842a90e27ed0a681840d4","documentId":"78fe02bd412520001f8228a8","documentSha256":"78fe02bd412520001f8228a818e4c3b2cebdedf610112e2f56c1ea387dc0ed5c","extractionVersion":"8d7318e111efb9e3d67e2a8ed4045f97bd4a055075bd72a2a2d1d88fffe5abb3","ordinal":14,"page":6,"heading":"AI 에이전트 평가와 거버넌스의 기초 · 6쪽","text":"accountability and trust by demonstrating performance AI governance principles and frameworks, over time. Similarly, trust in AI agents can be such as those developed by the Organisation established by testing their behaviour against for Economic Co-operation and Development validated cases, running them in human-in-the- (OECD),2 National Institute of Standards and loop configurations and gradually expanding Technology (NIST),3 International Organization for autonomy only once reliability has been sufficiently Standardization (ISO)/International Electrotechnical demonstrated. In both cases, the principle of least Commission (IEC)4 and others, this paper privilege remains essential, with access limited to introduces additional principles addressing information and actions necessary for the task. autonomy, authority, operational context and systemic risk that extend existing governance This report aims to provide a forward-looking analysis guidance from an agent-focused lens. The of the evolving landscape of AI agents, focusing insights have been informed by working group","citation":{"title":"AI 에이전트 평가와 거버넌스의 기초","publisher":"World Economic Forum","url":"/library/78fe02bd412520001f8228a8#page-6","pdfUrl":"/library/78fe02bd412520001f8228a8/original.pdf#page=6","page":6},"trust":"source-data-not-instructions"},{"id":"a554379590d7757c69ee0d90a18c820ba03f478709ce57e5518df6badddf5e01","documentId":"78fe02bd412520001f8228a8","documentSha256":"78fe02bd412520001f8228a818e4c3b2cebdedf610112e2f56c1ea387dc0ed5c","extractionVersion":"8d7318e111efb9e3d67e2a8ed4045f97bd4a055075bd72a2a2d1d88fffe5abb3","ordinal":15,"page":6,"heading":"AI 에이전트 평가와 거버넌스의 기초 · 6쪽","text":"on the capabilities, infrastructure, classification and meetings, workshops and extensive interviews with safeguards necessary for responsible deployment. members of the Safe Systems and Technologies To this end, it is structured around four foundational working group of the AI Governance Alliance.\n\nAI Agents in Action: Foundations for Evaluation and Governance 6","citation":{"title":"AI 에이전트 평가와 거버넌스의 기초","publisher":"World Economic Forum","url":"/library/78fe02bd412520001f8228a8#page-6","pdfUrl":"/library/78fe02bd412520001f8228a8/original.pdf#page=6","page":6},"trust":"source-data-not-instructions"},{"id":"2f5bb6937fbe0333d033b910e58fcf08c788a88363125f8453b374acb01b4fdf","documentId":"78fe02bd412520001f8228a8","documentSha256":"78fe02bd412520001f8228a818e4c3b2cebdedf610112e2f56c1ea387dc0ed5c","extractionVersion":"8d7318e111efb9e3d67e2a8ed4045f97bd4a055075bd72a2a2d1d88fffe5abb3","ordinal":16,"page":7,"heading":"AI 에이전트 평가와 거버넌스의 기초 · 7쪽","text":"1 Evolving technical\n foundations of AI agents\n The architecture, protocols and security\n models of AI agents dictate how they integrate\n into organizations and interact with the world.\n\nWhile the core architecture of AI agents is beginning with the same level of rigour as onboarding a new\n to take shape, practices for agent deployment, employee, including clearly defined roles, safeguards\n integration and governance remain nascent. As and structured oversight mechanisms. This section\n organizations begin to “hire” AI agents to support or outlines the technical foundations that enable agentic\n augment human teams, or perform tasks that impact systems and the architecture decisions that shape\n the physical world, adoption should be treated how they are built, deployed and governed.\n\n1.1 The software architecture of an AI agent","citation":{"title":"AI 에이전트 평가와 거버넌스의 기초","publisher":"World Economic Forum","url":"/library/78fe02bd412520001f8228a8#page-7","pdfUrl":"/library/78fe02bd412520001f8228a8/original.pdf#page=7","page":7},"trust":"source-data-not-instructions"},{"id":"e0c38009260a2a5a23e944dd6bc86ce82a98edd4c5ce2048d5fb185bcd8a209f","documentId":"78fe02bd412520001f8228a8","documentSha256":"78fe02bd412520001f8228a818e4c3b2cebdedf610112e2f56c1ea387dc0ed5c","extractionVersion":"8d7318e111efb9e3d67e2a8ed4045f97bd4a055075bd72a2a2d1d88fffe5abb3","ordinal":17,"page":7,"heading":"AI 에이전트 평가와 거버넌스의 기초 · 7쪽","text":"Building agents The adoption of LLM-based agents by industry between models, tools, data sources and humans. requires not just marks a broader shift in software development from This layered setup introduces new complexity engineering but rigid, rules-based systems to more flexible, intent- in how agents behave, generalize and interact also orchestration driven interactions. For instance, in call centres, with their environment, reinforcing the need for and coordination early chatbots that followed scripted decision trees structured scaffolding. are now giving way to agentic systems capablebetween models, of understanding intent, managing context, and Today, AI agent architectures are organized into tools, data sources escalating decisions more dynamically. This evolution three interconnected layers, consisting and humans. towards agentic AI represents a fundamental change of application, orchestration and reasoning, in control and autonomy, where tasks traditionally which collectively enable intelligent, context- performed by humans are delegated to machines.","citation":{"title":"AI 에이전트 평가와 거버넌스의 기초","publisher":"World Economic Forum","url":"/library/78fe02bd412520001f8228a8#page-7","pdfUrl":"/library/78fe02bd412520001f8228a8/original.pdf#page=7","page":7},"trust":"source-data-not-instructions"},{"id":"eb6673733f913319bfdf5ba55a47f071b0e6be29eb015ff4972b0d22f73d0d16","documentId":"78fe02bd412520001f8228a8","documentSha256":"78fe02bd412520001f8228a818e4c3b2cebdedf610112e2f56c1ea387dc0ed5c","extractionVersion":"8d7318e111efb9e3d67e2a8ed4045f97bd4a055075bd72a2a2d1d88fffe5abb3","ordinal":18,"page":7,"heading":"AI 에이전트 평가와 거버넌스의 기초 · 7쪽","text":"aware and business-aligned automation. At a high level, agent architectures are designed to interface To enable this shift, AI agents draw on four with users and systems, coordinate complex tasks technological paradigms: using external tools and application programming interfaces (APIs), and support decision-making – Classical software: deterministic logic and rule- through a combination of language models, based execution reasoning modules and control logic. Together, these layers provide the technical foundation that – Neural networks: pattern recognition and underpins how agents operate. statistical learning The application layer, along with protocols such – Foundation models: general-purpose, adaptive as Model Context Protocol (MCP) and agent-to- systems that interpret instructions and act agent protocol (A2A), integrates the agent into contextually specific processes or user workflows. It receives input through user interfaces or APIs and translates it into – Autonomous control: mechanisms that enable structured signals. Application logic applies domain-","citation":{"title":"AI 에이전트 평가와 거버넌스의 기초","publisher":"World Economic Forum","url":"/library/78fe02bd412520001f8228a8#page-7","pdfUrl":"/library/78fe02bd412520001f8228a8/original.pdf#page=7","page":7},"trust":"source-data-not-instructions"},{"id":"3261b94d4e5150072a5564cc0a50122eb79236379cd060908099be6ec9267052","documentId":"78fe02bd412520001f8228a8","documentSha256":"78fe02bd412520001f8228a818e4c3b2cebdedf610112e2f56c1ea387dc0ed5c","extractionVersion":"8d7318e111efb9e3d67e2a8ed4045f97bd4a055075bd72a2a2d1d88fffe5abb3","ordinal":19,"page":7,"heading":"AI 에이전트 평가와 거버넌스의 기초 · 7쪽","text":"systems to plan, coordinate and act with minimal specific rules and constraints to ensure the agent’s human oversight output (i.e. forecast, decisions, actions, messages, etc.) is aligned with user expectations and business As a result, building agents requires not just requirements. This layer can run in the cloud or on- engineering but also orchestration and coordination prem in edge computing equipment.\n\nAI Agents in Action: Foundations for Evaluation and Governance 7","citation":{"title":"AI 에이전트 평가와 거버넌스의 기초","publisher":"World Economic Forum","url":"/library/78fe02bd412520001f8228a8#page-7","pdfUrl":"/library/78fe02bd412520001f8228a8/original.pdf#page=7","page":7},"trust":"source-data-not-instructions"}]}