《世界經濟論壇&凱捷：2024駕馭人工智能前沿：人工智能代理(AI Agent)的演進和影響白皮書（英文版）（28頁）.pdf》由會員分享，可在線閱讀，更多相關《世界經濟論壇&凱捷：2024駕馭人工智能前沿：人工智能代理(AI Agent)的演進和影響白皮書（英文版）（28頁）.pdf（28頁珍藏版）》請在三個皮匠報告上搜索。

1、Navigating the AI Frontier:A Primer on the Evolution and Impact of AI AgentsW H I T E P A P E RD E C E M B E R 2 0 2 4In collaboration with CapgeminiImages:Getty ImagesDisclaimer This document is published by the World Economic Forum as a contribution to a project,insight area or interaction.The fin

2、dings,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.2024

3、 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 3Executive summary 4Introduction 51 Definition of an AI agent 62

4、 The evolution of AI agents 82.1.Key technological trends 92.2.Types of AI agents 102.3.Advanced AI agents 122.4.AI agent system 132.5.The future of AI agents:Towards multi-agent systems 143 Looking ahead 173.1.Key benefits 183.2.Examples of risks and challenges 183.3.Addressing the risk and challen

5、ges 20Conclusion 22Contributors 23Endnotes 26Navigating the AI Frontier:A Primer on the Evolution and Impact of AI Agents2ForewordFernando Alvarez Chief Strategy and Development Officer,CapgeminiJeremy Jurgens Managing Director,World Economic Forum Navigating the AI Frontier:A Primer on the Evolutio

6、n and Impact of AI AgentsDecember 2024In the contemporary world,where technology is rapidly reshaping every aspect of our lives,AI agents are emerging as transformative tools that are redefining human interactions and the operation of our society.These agents,which began as simple computer programs,

7、have evolved into sophisticated systems with the capability for autonomous decision-making.This evolution signifies a major shift,positioning AI agents as active participants in crucial sectors such as healthcare,education,financial services and beyond.The advancement of AI agents brings with it a w

8、ealth of exciting possibilities and transformative potential.Their ability to manage complex tasks with minimal human intervention offers the promise of significantly increased efficiency and productivity.However,as we step into this AI-driven era,it is essential to not only harness the immense bene

9、fits these technologies offer,but also to address the challenges they present.Issues such as ethical considerations require careful attention and proactive management.Ensuring that AI development aligns with societal values and aspirations is paramount to its successful integration into daily life.T

10、he aim of these innovations is to amplify human ingenuity not to replace it within our economy.This comprehensive overview serves as an important resource for those involved in shaping the future of AI technology.By exploring the capabilities and implications of AI agents,stakeholders can better und

11、erstand how to leverage the power of these systems to drive meaningful progress across various sectors.It is through this understanding that we can ensure AI technologies are developed responsibly and used in ways that enhance human well-being.With careful stewardship,AI agents can become invaluable

12、 allies in fostering innovation and improving quality of life worldwide.In partnership,the World Economic Forum and Capgemini have joined forces through the AI Governance Alliance to advance this critical topic in collaboration with the AI community.Navigating the AI Frontier:A Primer on the Evoluti

13、on and Impact of AI Agents3Executive summaryThis paper examines the development and functionality of AI agents and the implications of their use amid rapid advances in large language and multimodal models.Defined as autonomous systems that sense and act upon their environment to achieve goals,artifi

14、cial intelligence(AI)agents are being deployed in a wide range of roles in different industries.This requires the adaptation of governance frameworks to ensure responsible adoption.AI agents,comprising components such as sensors and effectors,have evolved from rule-based systems to advanced models c

15、apable of complex decision-making and independent operation.Enabled by breakthroughs in deep learning,reinforcement learning and the transformer architecture,1 AI agents span applications from workflow automation to personal assistants.This progression now encompasses more sophisticated utility-base

16、d AI agents that incorporate memory,planning and tool integration,broadening their capabilities and relevance.The benefits of AI agents include productivity gains,specialized support and improved efficiency in sectors such as healthcare,customer service and education.However,AI agents also present n

17、ovel risks,including potential misalignment,2 along with ethical concerns about transparency and accountability.Future advances in the area are likely to involve multi-agent systems(MAS),where AI agents collaborate to address complex challenges such as urban traffic management.More advanced systems

18、introduce new demands for interoperability and communication standards to function effectively,while these protocols still need to be debated and agreed upon by a wider community.This paper highlights the need for robust governance,ethical guidelines and a cross-sectoral consensus to integrate AI ag

19、ents safely into society.As more advanced AI agents continue to proliferate,it is imperative that their transformative potential remains balanced with essential safety,security and governance considerations.Navigating the AI Frontier:A Primer on the Evolution and Impact of AI Agents4IntroductionAI a

20、gents are becoming more advanced,with significant implications for decision-making,accountability and oversight.As artificial intelligence(AI)continues to advance and integrate into various sectors of the economy and society,understanding the role of AI agents,their capabilities and likely impact is

21、 crucial for business leaders,policy-makers and other stakeholders involved in shaping the future of AI development,implementation and governance.The concept of an agent an entity that perceives its environment through sensors and acts on it through effectors has been constantly evolving since the b

22、eginning of task automation.3 With recent advances in large language models(LLM AI models that process natural language)and large multimodal models(LMM AI models that process natural language,images,video and/or audio),the concept of AI agents is moving into a new phase of rapid development and expe

23、rimentation.This phase is currently seeing the emergence of a range of novel use cases from coding assistants to workflow automation,personal assistants and many more areas of application.As AI agents continue to advance,society is gradually progressing towards the development of innovative systems

24、with increased autonomy,capable of completing tasks with minimal human involvement or guidance.This heralds a new era of AI-driven innovation and efficiency with the potential to affect every sector of the global economy.Given this far-reaching prospect,it is crucial to consider safety and governanc

25、e measures to guide the responsible development and implementation of advanced AI agents.4This paper first defines the concept of AI agents before outlining different types of agents and their evolution over time.The last section looks ahead and summarizes examples of emerging technical and the soci

26、oeconomic implications of deploying AI agents along with possible measures to mitigate risks.Navigating the AI Frontier:A Primer on the Evolution and Impact of AI Agents5Definition of an AI agent1An AI agent responds autonomously to inputs and its reading of its environment to make complex decisions

27、 and change the environment.Navigating the AI Frontier:A Primer on the Evolution and Impact of AI Agents6Based on the definition of the International Organization for Standardization,5 an AI agent can be broadly defined as an entity that senses percepts(sound,text,image,pressure etc.)using sensors a

28、nd responds(using effectors)to its environment.AI agents generally have the autonomy(defined as the ability to operate independently and make decisions without constant human intervention)and authority(defined as the granted permissions and access rights to perform specific actions within defined bo

29、undaries)to take actions to achieve a set of specified goals,thereby modifying their environment.The core components of an AI agent FIGURE 1:Figure 1 highlights how an agent is made up of several core components,including:User input:the external(e.g.human,another agent)input that the AI agent receiv

30、es.This could be instructions such as typing via a chat-based interface,voice-based commands or pre-recorded data.Environment:the bounds in which the AI agent operates.It serves as the area in which the agent applies its sensors and effectors to percept and modify its surroundings based on the input

31、s received and the actions decided upon by the control centre.The environment can be physical infrastructure such as the mapped area of an autonomous vehicle or digital infrastructure such as the intranet of a business for a coding agent.Sensors:mechanisms through which the agent perceives its envir

32、onment.Sensors can range from physical devices(e.g.cameras or microphones)to digital ones(e.g.queries to databases or web services).Control centre:typically makes up the core of the AI agent along with the model,such as an LLM.The control centre helps process information,make decisions and plan acti

33、ons.Based on the capabilities of the AI agent,the control centre involves complex algorithms and models that allow the agent to evaluate different options and choose the best course of action.Percepts:the data inputs that the AI agent receives about its environment,which could come from various sens

34、ors or other data sources.They represent the agents perception or understanding of its environment.Effectors:the tools an agent uses to take actions upon its environment.In physical environments,effectors might include robotic arms or wheels,while in the digital environment,they could be commands se

35、nt to other software systems,such as generating a data visualization or executing a workflow.Actions:represent the alterations made by effectors.In physical environments,actions might be pushing an object,whereas in digital environments they could be linked to updating a database.AI agentPerceptsEnv

36、ironmentActionsSensorsControl centreEffectorsDigitalinfrastructureUser inputPhysicalinfrastructureSource:World Economic ForumNavigating the AI Frontier:A Primer on the Evolution and Impact of AI Agents7The evolution of AI agents2Developers have transformed AI from rule-based systems to active agents

37、 capable of learning and adapting while engaged in a task.Navigating the AI Frontier:A Primer on the Evolution and Impact of AI Agents8The development of AI agents began in the 1950s,6 and since then they have evolved from simple rule-based systems to sophisticated autonomous entities capable of com

38、plex decision-making.Early AI was characterized by deterministic behaviour,relying on fixed rules and logic that made these systems predictable but unable to learn or adapt from new experiences.Advances in AI research introduced systems that could handle larger datasets and manage uncertainty,leadin

39、g to probabilistic outcomes and non-deterministic behaviour.This shift enabled more flexible and dynamic decision-making,moving beyond rigid frameworks.The 1990s marked a significant turning point,as machine learning applications became more widespread.AI systems began to learn from data,adapt over

40、time and improve performance.The introduction of neural networks during this period laid the foundation for deep learning,which has since become essential to modern AI.Since 2017,the rise of LLMs has transformed AIs capabilities in natural language understanding and generation.These models use vast

41、amounts of data to produce human-like text and engage in complex language-based tasks.Todays AI agents use various learning techniques,including reinforcement learning,or transfer learning,allowing them to continuously refine their abilities,adapt to new environments and make more informed decisions

42、.Key technological trends2.1Over the past 25 years,the increase in computing capacity,the availability of large quantities of data on the internet and novel algorithmic breakthroughs have enabled significant developments in the base technologies behind recent advances in the capabilities of AI agent

43、s.These are briefly described below.Large models Large language models(LLM)and large multimodal models(LMM)have revolutionized the capabilities of AI agents,particularly in natural language processing and the generation of text,image,audio and video.The emergence of large models has been driven by s

44、everal technological advances and by the transformer architecture,which has paved the way for a deeper understanding of context and word relationships,considerably improving the efficiency and performance of natural language processing tasks.7 In summary,advanced AI models have enabled better unders

45、tanding,generation and engagement with natural language.Machine learning and deep learning techniquesA range of techniques have greatly improved AI models through increased efficiency and greater specialization.Some examples of machine-and deep-learning techniques include:1.Supervised learning:facil

46、itates learning from labelled datasets,so the model can accurately predict or classify new,previously unseen data.82.Reinforcement learning:enables agents to learn optimal behaviours through trial and error in dynamic environments.Agents can continuously update their knowledge base without needing p

47、eriodic retraining.93.Reinforcement learning with human feedback:enables agents to adapt and improve through human feedback,specifically focusing on aligning AI behaviour with human values and preferences.104.Transfer learning:involves taking a pretrained model,typically trained on a large dataset(e

48、.g.to recognize cars)and adapting it to a new but related problem(e.g.to recognize trucks).115.Fine-tuning:involves taking a pretrained model and further training it on a smaller,task-specific dataset.This process allows the model to retain its foundational knowledge while improving its performance

49、on specialized tasks.12These and other learning paradigms are often used in combination and have dramatically expanded the problem-solving capabilities of AI agents in various areas of application.The evolution of AI agents is detailed in Figure 2,while the agent types are further expanded in the fo

50、llowing section.Navigating the AI Frontier:A Primer on the Evolution and Impact of AI Agents9Evolution of AI agents capabilities FIGURE 2:TABLE 1:Technology trendDeterministicNon-deterministicCurrent stateAgent typeSimple reflexCondition action rulesInternal model ofthe environmentTransfer learning

51、and reinforcement learningEvaluating scenariosto choose the bestoutcomeCollaborativemethodologies thatrepresent the currentstate of the artBasic anti-virus softwareSmartthermostatAdvancedchess AIAutonomous drivingSmart citytraffic plannerModel-basedGoal-basedUtility-basedMachine learning and deep le

52、arning techniquesLarge modelsMulti-agent systemsFuture typeAgent examplesKey characteristicsSource:World Economic ForumTypes of AI agents2.2This section outlines different types of AI agent and traces their evolution,highlighting the key technological advances that have supported their development.A

53、I agents can be considered as either deterministic or non-deterministic,based on their defining characteristics,which are outlined below.Defining characteristics of deterministic and non-deterministic AI agentsDeterministic AI agentsNon-deterministic AI agentsRule-based:operate with fixed rules and

54、logic,meaning the same input will always produce the same output.Data-driven and probabilistic:make decisions based on statistical patterns in data,with outcomes that are not fixed but instead are probabilistic.Predictable behaviour:the decision-making process is transparent and consistent,which mak

55、es the outcomes predictable.Flexible and adaptive:able to learn from data,adapt to new situations and handle uncertainty,often resulting in varied outcomes for similar inputs.Limited adaptability:these systems cannot learn from new data or adjust to changes;they follow only predefined paths.Complex

56、decision-making:use algorithms that factor in probabilities,randomness or other non-deterministic elements,allowing for more nuanced and complex behaviours.Navigating the AI Frontier:A Primer on the Evolution and Impact of AI Agents10Simple reflex agents operate based on a perception of their enviro

57、nment,without consideration of past experiences.13 Instead,they follow predefined rules to map specific inputs to specific actions.The implementation of conditionaction rules allows for rapid responses to environmental stimuli.These early agents are simple rule-based machines or algorithms designed

58、to provide static information and unable to adapt or change course.Basic spam filters using keyword matching Simple chatbots with predefined responses Automated email responders that send prewritten replies following specific triggersModel-based reflex agents are designed to track parts of their env

59、ironment that are not immediately visible to them.14 They do this by using stored information from previous observations,allowing them to make decisions based on both current inputs and past experiences.By basing their actions on both current perceptions and their internal model,these agents are mor

60、e adaptable than simple reflex agents even though they are also governed by conditionaction rules.Simple reflex agentsModel-based reflex agentsGoal-based agentsUtility-based agents Smart thermostats that optimize energy usage by adjusting to current and historical temperature data,as well as user pr

61、eferences Smart robotic vacuum cleaners that use sensors and maps to navigate efficiently,avoiding obstacles and optimizing cleaning paths Modern irrigation systems that use sensors to collect real-time data on environmental factors such as soil,moisture,temperature and precipitation,to optimize wat

62、er dispensationGoal-based agents are able to take future scenarios into account.This type of agent considers the desirability of actions outcomes and plans to achieve specific goals.15 The integration of goal-oriented planning algorithms allows the agent to make decisions based on future outcomes,ma

63、king them suitable for complex decision-making tasks.Advanced chess AI engines that have the goal of winning the game,planning moves that maximize the probability of success and considering a long-term strategy Route optimization systems for logistics that set goals for efficient delivery and plan o

64、ptimal routes by setting clear priorities Customer service chatbots that set goals to resolve customer issues and plan conversation flows to achieve their goals efficientlyUtility-based agents employ search and planning algorithms to tackle intricate tasks that lack a straightforward outcome,thereby

65、 going beyond simple goal achievement.They use utility functions to assign a weighted score to each potential state,facilitating optimal decision-making in scenarios with conflicting goals or uncertainty.Rooted in decision theory,this method allows for more advanced decision-making in complex enviro

66、nments.These agents can balance multiple,possibly conflicting objectives according to their relative significance.16 Autonomous driving systems that optimize safety,efficiency and comfort while evaluating trade-offs such as speed,fuel efficiency and passenger comfort Portfolio management systems suc

67、h as robot-advisers that make financial decisions based on utility functions that weigh risk,return and client preferences Healthcare diagnosis assistants that analyse patient medical records,label patient data (e.g.tumour detection)and optimize treatment strategy recommendations in cooperation with

68、 doctorsTypeDefinitionExamplesNavigating the AI Frontier:A Primer on the Evolution and Impact of AI Agents11Advanced AI agents2.3The architecture of many current AI agents is often based on or linked to LLMs,which are configured in complex ways.Figure 3 presents a simplified overview of the key comp

69、onents leading to current breakthroughs in AI agents and their growing range of capabilities.Key components of advanced AI agents FIGURE 3:The AI agent begins with user input,which is directed to the agents control centre.The user input could be a prompt given to carry out an instruction.The control

70、 centre directs the user input to the model,which forms the core algorithmic foundation of the AI agent.This model could be an LLM or an LMM,depending on the applications needs.The model then processes the input data from the users instructions to generate the desired result.17At the core of the arc

71、hitecture is the control centre,a crucial component that manages the flow of information and commands throughout the system.It acts as the orchestration layer,directing inputs to the model and routing the output to appropriate tools or effectors.In simple terms,this layer orchestrates the flow of in

72、formation between 1)user inputs,2)decision-making and planning,3)memory management,4)access to tools and 5)the effectors of the system enabling action in digital or physical environments.18 The decision-making and planning component of an AI agent uses the models outputs to assist in decision-making

73、 and planning of multistep processes.In this segment,advanced features such as chain-of-thought(CoT)reasoning are implemented,which allows the AI agent to engage in multistep reasoning and planning.CoT is a technique where an AI agent systematically processes and articulates intermediate steps to re

74、ach a conclusion,which enhances the agents ability to solve complex problems in a transparent manner,as each step of the models underlying reasoning is reproduced in natural language.19Memory management is vital for the continuity and relevance of operations.This component ensures that the AI agent

75、remembers previous interactions and maintains context.This is essential for tasks that require historical data to inform decisions or for maintaining conversational context in chatbots.Tools enable the AI agent to access and interact with multiple functions or modalities.For example,in an online set

76、ting,an AI agent could have access to external tools such as web searches to gather real-time information and scheduling tools to manage appointments and send reminders,as well as project management software to track tasks and deadlines.In terms of modalities,an AI agent could use natural language p

77、rocessing tools alongside image recognition capabilities to perform tasks that require understanding of text-based as well as visual-based data sources.Once decisions are made or plans set,the effectors component of the AI agent executes the required actions.This could involve interacting AI agentPe

78、rceptsEnvironmentActionsSensorsLearningDigitalinfrastructureUser inputPhysicalinfrastructureEffectorsControl centreModelDecision-making andplanningMemorymanagementToolsSource:World Economic ForumNavigating the AI Frontier:A Primer on the Evolution and Impact of AI Agents12with the physical world(in

79、robotics),executing a software function or providing recommendations and decisions to human users.The learning component is intrinsic to the model and enables the AI agent to improve its performance over time as the model gathers more input,using machine learning and deep learning techniques as ment

80、ioned in section 2.1.The application layer surrounds the control centre,models and other components,acting as the interface between the AI agent and its environment.It interprets the outputs from the control centre and adapts them to specific tasks or domains.For example,in a healthcare AI agent,the

81、 application layer would translate model outputs into diagnostics,treatment recommendations or medical alerts through an appropriate user interface.In summary,when the varying components of an advanced AI agent come together,they represent the agents ability to model the environment,maintain memory

82、or knowledge storage with beliefs and preferences,as well as inherent abilities to learn,plan,make decisions,perceive(sense),act(interact)and communicate with the agents surroundings.Example of an advanced AI agent:AI agent infotainment systemAn AI agent in a cars infotainment system acts as a smart

83、 assistant,activated through voice commands to manage navigation,entertainment,climate controls and other vehicle settings.It processes live traffic,weather and driver preferences to optimize routes,suggesting alternatives around delays or hazards.The agent personalizes entertainment based on user h

84、abits,recommends nearby stops such as restaurants or fuel stations and proactively provides updates such as low fuel alerts or optimal recharging points for electric vehicles all while ensuring the driver remains focused on the road.AI agent system2.4An AI agent system is an organized structure that

85、 integrates multiple heterogeneous(e.g.rule-and goal-based agents)or homogeneous(e.g.goal-based only)AI agents.20 Each agent is typically specialized,possessing its own capabilities,knowledge and decision-making processes,while sharing data to collaboratively achieve the goal of the system.Several d

86、esigns are possible,such as:Mixture-of-agents,where each agent is called sequentially,with agents processing the outputs from each previous agent21 Central orchestration,which coordinates calls of agents and manages the inputs and outputs accordinglyThe AI agent system is designed to ensure that eac

87、h agent contributes to the overall objective,whether it involves managing complex real-time processes such as autonomous driving,optimizing industrial processes or coordinating activities;for example,in smart city infrastructure.By dividing the workload among specialized agents,the system can handle

88、 dynamic environments and adapt to changing conditions,ensuring optimal performance.Example of an AI agent system:Autonomous vehicle AI agent systemA human user gets into an autonomous vehicle(AV).The AV is comprised of an AI agent system that includes agents for perception,path planning,localizatio

89、n for finding its specific place on the road and control to steer and brake.The perception and localization agents are dedicated to continuously mapping the environment through sensors,the global positioning system(GPS)and cameras.The planning agent calculates the optimal trajectory by factoring in

90、real-time traffic,weather and road conditions.The control agent handles the vehicles core mechanics,such as braking,accelerating and steering.22 The AI agent infotainment system serves as the interface with the passenger,and handles elements such as processing voice commands and adjusting routes,cli

91、mate,entertainment or other in-car settings based on user preferences.23 All agents work together in a coordinated and centralized manner to ensure the vehicle reaches its destination safely and efficiently,prioritizing both passenger comfort and safety.241.2.Navigating the AI Frontier:A Primer on t

92、he Evolution and Impact of AI Agents13Network architectureSupervised architectureAI agentAI agentsuperivsorAI agentsystemAI agentsystemAI agentAI agentAI agentsystemAI agentsystemThe future of AI agents:Towards multi-agent systems2.5Multi-agent systems(MAS)consist of multiple independent AI agents a

93、s well as AI agent systems that collaborate,compete or negotiate to achieve collective tasks and goals.25 These agents can be autonomous entities,such as software programs or robots,each typically specialized with its own set of capabilities,knowledge and decision-making processes.This allows agents

94、 to perform tasks in parallel,communicate with one another and adapt to changes in complex environments.The architecture of a MAS is determined by the desired outcomes and the goals of each participating agent or system.There are several architectural types,26 for example:Network architecture:In thi

95、s set-up,all agents or systems can communicate with one another to reach a consensus that aligns with the MASs objectives.For example,when autonomous vehicles(AVs)park in a tight space,they communicate to avoid collision.In this case,the MAS objective to prevent accidents aligns with each AVs goal o

96、f safe navigation,allowing them to coordinate effectively and reach consensus.Supervised architecture:In this model,a“supervisor”agent coordinates interactions among other agents.It is useful when agents goals diverge,and consensus may be unattainable.The supervisor can mediate and prioritize the MA

97、Ss objectives while considering each agents unique goals,thereby finding a compromise.An example could be when a buyer and seller agent cannot reach agreement on a transaction,which is then mediated by an AI agent supervisor.Examples of MAS architecture FIGURE 4:While current efforts largely focus o

98、n developing AI agents within closed environments or specific software ecosystems,the future is likely to see multiple agents collaborating in different domains and applications.In MAS,different types of agent could work together to tackle increasingly complex tasks that require multistep processes,

99、integrating expertise from various fields to achieve more sophisticated outcomes.These agents can communicate and interact within a broader adaptive system,enabling them to handle both specific tasks and complex situations more efficiently than a single agent,or even an AI agent system,could on its

100、own.In some cases,multi-agent systems address the limitations of single-agent systems,such as scalability issues,lack of resilience in the event of Source:World Economic ForumNavigating the AI Frontier:A Primer on the Evolution and Impact of AI Agents14failure or errors and limited skill sets.By dis

101、tributing tasks among multiple agents,MAS could increase both efficiency and capability.In theory,multi-agent systems are highly adaptable,as agents can be dynamically added or removed,allowing the system to respond to changing environments and requirements.This scalability is crucial for applicatio

102、ns that need to grow or evolve over time without extensive re-engineering.In many ways,multi-agent systems can be considered as a future type of system that could coordinate agent actions among multiple users or organizations through human-comprehensible language or to-be-determined AI agent protoco

103、ls.FIGURE 5:The structure and relationships among the AI agent,AI agent system and multi-agent systemAdvanced AI agentEXAMPLE:AI agent infotainment systemEXAMPLE:Fully autonomous vehicleEXAMPLE:Connected smart city coordinationAI agent systemMulti-agent systemSensorsLearningAI agent 1OrchestrationAI

104、 agent 3Other AI agentAI agent 2AI agent system 1AI agent 2AI agent 3Other agentsEffectorsControl centreModelDecision-making and planningMemorymanagementTools1.2.3.Example of a multi-agent system:Smart city traffic management with vehicle-to-everything(V2X)communication Source:World Economic ForumIn

105、 a smart city,a multi-agent system(MAS)manages traffic flow in real time,using vehicle-to-everything(V2X)communication,enabling vehicles to interact with other vehicles,pedestrians and road infrastructure.27 Each traffic signal is controlled by an AI agent system that communicates with nearby signal

106、s,public transport systems,emergency services and parking services to check availability.Vehicles,equipped with their own AI agent system,share data such as speed,location and road conditions,allowing for coordinated actions to enhance road safety,traffic efficiency and energy usage.For example,if a

107、n accident occurs,AI agents can reroute traffic,adjust signal timings,notify emergency services and communicate with vehicles and pedestrians to avoid the area,all with minimal human intervention.This system optimizes traffic flow,improves road safety and reduces energy consumption by dynamically ad

108、apting to real-time conditions.For instance,if a parking lot is full,the system can direct vehicles to available parking further away,even if it conflicts with the drivers and the onboard AI agents preference for proximity.3.Navigating the AI Frontier:A Primer on the Evolution and Impact of AI Agent

109、s15Interoperability of multi-agent systemsOne technical challenge in multi-agent systems is associated with enabling effective communication between different AI agents and AI agent systems.28 In some cases,interactions are limited by the boundaries of native application environments,restricting the

110、 potential of AI agents to narrower and more specialized subdomains,where control is more easily retained.The interoperability of AI agents relies on common communication protocols,which are the rules and standards governing how AI agents exchange information.These protocols can generally be categor

111、ized in two types:Predefined protocols:these are based on established agent communication languages and ontologies.Since they are predefined,the communication patterns are predictable and consistent;however,they may not adapt well to dynamic environments where new communication needs arise.29 Emerge

112、nt protocols:these allow agents to learn how to communicate effectively based on their experiences,often using reinforcement learning techniques.This enables agents to adapt their communication strategies to changing environments and tasks.30 However,decoding and understanding emergent communication

113、 remains an ongoing research challenge.31A good understanding of the messages exchanged between AI agents is essential,otherwise it could affect the overall reliability of multi-agent systems.This inconsistency could lead to misunderstandings or misaligned actions when agents collaborate,especially

114、in complex environments requiring precise coordination.To enhance the transparency of multi-agent interaction,the information exchanged needs to be easily accessible and interpretable by humans.Navigating the AI Frontier:A Primer on the Evolution and Impact of AI Agents16Looking ahead3AI agents have

115、 the potential to tackle challenging tasks with great efficiency.But they carry associated risks such as malfunction,malicious use and unwanted socioeconomic effects.Navigating the AI Frontier:A Primer on the Evolution and Impact of AI Agents17Key benefits3.1By scaffolding capabilities such as reaso

116、ning,planning and self-checking on top of LLMs,more capable AI agents emerge that hold the potential to dramatically increase users productivity and absolve them from certain tasks.This could involve completing tasks beyond users skill sets,such as specialized coding,or partially or fully offloading

117、 tedious tasks that can be done more cheaply,quickly and at a greater scale than before.Additionally,the application of AI agents can play a crucial role in addressing the shortfall of skills in various industries,filling the gaps in areas where human expertise is lacking or in high demand.Key chara

118、cteristics of greater autonomy increasingly allow AI agents to tackle open-ended,real-world challenges that at one time were beyond them for example,helping in scientific discovery,improving the efficiency of complex systems such as supply chains or electrical grids,managing rare and unusual scenari

119、os in processes that are too infrequent to justify traditional automation,or enabling physical robots that can manipulate objects and navigate physical environments.32 Software development AI agents can help generate,run and check code and other artefacts needed,allowing software developers to focus

120、 on higher value-added activities.Examples of the benefits of applications of AI agents include:Healthcare AI agents could improve diagnostics and personalized treatment,reducing hospital stays and costs through data analysis and decision-making support.For example,in under-resourced areas,AI agents

121、 could help alleviate the workload of clinical specialists by assisting doctors in developing tailored treatment plans.33Enhanced customer experience AI agent-based chatbots or virtual assistants can offer personalized,round-the-clock support,increasing customer satisfaction.They have the potential

122、to provide consistently accurate responses,helping businesses maintain communication quality and resolve customer issues efficiently.34Education AI agents could help personalize learning experiences by adapting content to each students needs,offering real-time feedback and supporting teachers with g

123、rading and administrative tasks.This allows educators to focus more on creative and interactive learning experiences.Finance AI agents could help enhance fraud detection,optimize trading strategies and offer personalized financial advice.They can analyse large datasets to identify patterns and trend

124、s,providing faster and more accurate insights for decision-making.Navigating the AI Frontier:A Primer on the Evolution and Impact of AI Agents18While AI agents have the potential to offer numerous benefits,they also come with inherent risks,as well as novel safety and security implications.For examp

125、le,an AI system independently pursuing misaligned objectives could cause immense harm,especially in scenarios where the AI agents level of autonomy increases while the level of human oversight decreases.AI agents learning to deceive human operators,pursuing power-seeking instrumental goals or collud

126、ing with other misaligned agents in unexpected ways could pose entirely novel risks.35Agent-specific risks can be both technical and normative.Challenges associated with AI agents stem from technical limitations,ethical concerns and broader societal impacts often associated with a systems level of a

127、utonomy and the overall potential of its use when humans are removed from the loop.Without a human in the loop at appropriate steps,agents may take multiple consequential actions in rapid succession,which could have significant consequences before a person notices what is happening.36AI agents can a

128、lso amplify known risks associated with the domain of AI and could introduce entirely new risks that can be broadly categorized into technical,socioeconomic and ethical risks.Technical risksExamples of technical risks include:Risks from malfunctions due to AI agent failures:AI agents can amplify the

129、 risks from malfunctions by introducing new classes of failure modes.LLMs,for example,can enable agents to produce highly plausible but incorrect outputs,presenting risks in ways that were not possible with earlier technologies.These emerging failure modes add to traditional issues such as inaccurat

130、e sensors or effectors and encompass capability-and goal-related failures,as well as increased security vulnerabilities that could lead to malfunctions.37 Capability failures occur when an AI agent fails to perform the tasks it was designed for,due to limitations in its ability to understand,process

131、 or execute the required actions.Goal-related failures occur when a system is highly capable but nevertheless pursues the wrong goal.These issues can be caused by:Specification gaming:When AI agents exploit loopholes or unintended shortcuts in their programming to achieve their objectives,rather tha

132、n fulfilling their goals.38 Goal misgeneralization:When AI agents apply their learned goals inappropriately to new or unforeseen situations.39 Deceptive alignment:When AI agents appear to be aligned with the intended goals during training or testing,but their internal objectives differ from what is

133、intended.40 Malicious use and security vulnerabilities:AI agents can amplify the risk of fraud and scams increasing both in volume and sophistication.More capable AI agents can facilitate the generation of scam content at greater speeds and scale than previously possible,and AI agents can facilitate

134、 the creation of more convincing and personalized scam content.For example,AI systems could help criminals evade security software by correcting language errors and improving the fluency of messages that might otherwise be caught by spam filters.41 More capable AI agents could automate complex end-t

135、o-end tasks that would lower the point of entry for engaging in harmful activities.Some forms of cyberattacks could,for example,be automated,allowing individuals with little domain knowledge or technical expertise to execute large-scale attacks.42 Challenges in validating and testing complex AI agen

136、ts:The lack of transparency and non-deterministic behaviour of some AI agents creates significant challenges for validation and verification.In safety-critical applications,this unpredictability complicates efforts to assure system safety,as it becomes difficult to demonstrate reliable performance i

137、n all scenarios.43 While failures in agent-based systems are expected,the varied ways in which they can fail adds further complexity to safety assurance.Failsafe mechanisms are essential but could be harder to design due to uncertainty on potential failure modes.44Socioeconomic risksExamples of soci

138、oeconomic risks include:Over-reliance and disempowerment:Increasing autonomy of AI agents could reduce human oversight and increase the reliance on AI agents to carry out complex tasks,even in high-stakes situations.Malfunctions of the AI agents due to design flaws or adversarial attacks may not be

139、immediately apparent if humans are not in the loop.Additionally,disabling an agent could be difficult if a user lacks the required expertise or domain knowledge.45 Pervasive interaction with intelligent AI agents could also have long-term impacts on individual 3.2 Examples of risks and challengesNav

140、igating the AI Frontier:A Primer on the Evolution and Impact of AI Agents19and collective cognitive capabilities.For example,increased reliance on AI agents for social interactions,such as virtual assistants,AI agent companions,therapists and so on could contribute to social isolation and possibly a

141、ffect mental well-being over time.Societal resistance:Resistance to the employment of AI agents could hamper their adoption in some sectors or use cases.Employment implications:The use of AI agents is likely to transform a variety of jobs by automating many tasks,increasing productivity and altering

142、 the skills required in the workforce,thus causing partial job displacement.Such displacement could primarily affect sectors reliant on routine and repetitive tasks,in industries such as manufacturing or administrative services.Financial implications:Organizations could face higher costs associated

143、with the deployment of AI agents,such as expenses for securing software systems against cyberthreats and managing associated operational risks.Ethical risksExamples of ethical risks include:Ethical dilemmas in AI decision-making:The autonomous nature of AI agents raises ethical questions about their

144、 decision-making capabilities in critical situations.Challenges in ensuring AI transparency and explainability:Many AI models operate as“black boxes”,making decisions based on complex and opaque processes,thereby making it difficult for users to understand or interpret how decisions are made.46 A la

145、ck of transparency could lead to concerns about potential errors or biases in the AI agents decision-making capabilities,which would hinder trust and raise issues of moral responsibility and legal accountability for decisions made by the AI agent.Addressing the risk and challenges3.3To enable the au

146、tonomy of AI agents for cases where it would greatly improve outcomes,several challenges must be addressed.These challenges include safety and security-related assurance,regulation,moral responsibility and legal accountability,data equity considerations,data governance and interoperability,skills,cu

147、lture and perceptions.47 Addressing these challenges requires a comprehensive approach throughout the stages of design,development,deployment and use of AI agents as well as changes across policy and regulation.As advanced AI agents and multi-agent systems continue to evolve and integrate into vario

148、us aspects of digital infrastructure,associated governance frameworks that take increasingly complex scenarios into consideration need to be established.In assessing and mitigating the risks of potential harm from AI agents,it is essential to understand the specific application and environment of th

149、e AI agent(including stakeholders that may be affected).The risks of potential harm from an AI agent stem largely from the context in which it is deployed.48 In high-stakes environments such as healthcare or autonomous driving,even small errors or biases can lead to significant consequences for the

150、users of such systems.Conversely,in low-stakes contexts,such as customer service,the same AI agent might pose minimal risks,as mistakes are less likely to cause serious harm.Within the context of a specific application and environment,it is important to adopt a risk analysis methodology that systema

151、tically identifies,categorizes and assesses all of the risks associated with the AI agent.Such an approach helps ensure that appropriate and effective mitigation mechanisms and strategies can be implemented by relevant stakeholders at the technical,socioeconomic and ethical levels.Technical risk mea

152、suresExamples of technical risk measures:Improving information transparency:Where,why,how,and by whom information is used is critical for understanding how a system operates and why certain decisions are made by the agent.Measures can be implemented to improve the transparency of AI agents such as t

153、he integration of behavioural monitoring and implementation of thresholds,triggers and alerts that involve continuous observation and analysis of the agents actions and decisions.Implementing behavioural monitoring helps to ensure that failures are better understood and properly mitigated when they

154、occur.49Navigating the AI Frontier:A Primer on the Evolution and Impact of AI Agents20Socioeconomic risk measuresExamples of socioeconomic risk measures:Public education and awareness:Developing and executing strategies to inform and engage the public are essential to mitigate the risks of over-reli

155、ance and disempowerment in social interactions with AI agents.These efforts should aim to equip individuals with a solid understanding of the capabilities and limitations of AI agents,allowing for more informed interactions,along with healthy integrations.A forum to collect public concerns:Acceptanc

156、e and involvement,trust and psychological safety are crucial to tackle societal resistance and for the proper adoption and integration of AI agents into various processes.Without sufficient human“buy-in”,the implementation of AI agents would face significant challenges.In addressing societal resista

157、nce and creating wider trust in AI agents and autonomous systems,it is important that public concerns are heard and addressed throughout the design and deployment of advanced AI agents.50 Thoughtful strategies for deployment:Organizations can embrace deliberate strategies around increased efficiency

158、 and task augmentation rather than focusing on outright worker replacement efforts.By prioritizing proactive measures such as retraining programmes,workers can be supported in transitioning to new or changed roles.Ethical risk measuresExamples of ethical risk measures:Clear ethical guidelines:Priori

159、tizing human rights,privacy and accountability are essential measures to ensure that AI agents make decisions that are aligned with human and societal values.51 Behavioural monitoring:Implementing measures that allow users to trace and understand the underlying reasoning behind an AI agents decision

160、s is necessary to mitigate transparency challenges.52 Behavioural monitoring can make system behaviour and decisions visible and interpretable,which enhances overall user understanding of interactions.This approach also strengthens the governance structure surrounding AI agents and helps increase st

161、akeholder accountability.53As the adoption of AI agents increases,critical trade-offs need to be made.Given the complex nature of many advanced AI agents,safety should be regarded as a critical factor alongside other considerations such as cost and performance,intellectual property,accuracy,and tran

162、sparency,as well as implied social trade-offs when it comes to deployment.The level of autonomy of advanced AI agents is likely to continue to increase due to ever more capable models and reasoning capabilities.54 The complexities of more advanced systems call for a multidisciplinary approach that i

163、ncludes diverse stakeholders,from scientists and researchers to psychologists,developers,system and service integrators,operators,maintainers,users and regulators,all of whom are needed to establish appropriate risk management frameworks and governance protocols for the deployment of more sophistica

164、ted AI agent systems.This white paper has taken a first step in outlining the landscape of frontier AI agents,but further research is needed to provide more details on the safety,security and socioeconomic implications as well as the novel governance measures required to address them.Navigating the

165、AI Frontier:A Primer on the Evolution and Impact of AI Agents21ConclusionAI agents are becoming more autonomous in their operation and decision-making,bringing potential benefits and risks.The development of AI agents has been marked by significant milestones,from the early days of simple reflex age

166、nts to sophisticated multi-agent systems.Recent advances in LLMs and LMMs have resulted in the next evolution of AI agents,which have moved from basic systems that react to immediate stimuli to complex entities capable of planning,learning and making decisions based on a comprehensive understanding

167、of their environment and user needs.The ongoing development of AI agents is fundamentally linked to increased autonomy,improved learning capabilities,enhanced decision-making abilities and multi-agent collaboration.As the architecture and emerging use cases for AI agents continue to proliferate,the

168、shift towards multi-agent systems that can collaborate in increasingly complex environments is likely to continue.Increased autonomy plays an important part in the evolution of AI agents and creates novel opportunities for new applications while also presenting unique risks to society.The introducti

169、on of AI agents will likely reduce the need for human involvement and oversight in some areas,bringing a more efficient approach to tedious tasks.However,a reduction in human oversight could also increase the risk of accidents.Furthermore,increased automation of workflows could be a way for maliciou

170、s actors to exploit novel vulnerabilities,while also exacerbating socioeconomic and ethical risks.The rapid advance of AI agent capabilities is set to be followed by a wave of innovation in AI agents,which could have the ability to transform the global economy and the roles of human labour in new an

171、d significant ways.Further research is necessary to explore the safety,security and societal impacts of AI agents and multi-agent systems,emphasizing both technical solutions and organizational governance frameworks.These efforts are critical for mitigating risks associated with the ongoing developm

172、ent,deployment and increasing use of more sophisticated AI agents in a range of domains.At this point,it is vital for stakeholders to come together throughout technical,civil society,applied and governance-facing communities to research,discuss and build consensus on novel governance mechanisms.This

173、 white paper has offered an initial exploration of the rapidly evolving landscape of AI agents,aiming to promote deeper understanding of this emerging field and spark conversation on responsible adoption and diffusion practices.Through equitable development,deployment and governance,the growing pres

174、ence of advanced AI agents holds the promise of driving positive societal transformation for many years to come.Navigating the AI Frontier:A Primer on the Evolution and Impact of AI Agents22ContributorsThe World Economic Forums AI Governance Alliance Safe Systems and Technologies working group conve

175、nes chief science officers and AI producers to advance thought leadership surrounding AI agents,from their architecture to applications,social implications,guardrails and governance structures.This initiative promotes the development of safety mechanisms and encourages collaboration on best practice

176、s for AI system design and implementation.55World Economic ForumBenjamin LarsenLead,Artificial Intelligence and Machine LearningCathy LiHead,AI,Data and Metaverse;Member of the Executive CommitteeStephanie TeeuwenSpecialist,Data Policy and AICapgeminiOlivier DentiData Architect,AI,Capgemini InventJa

177、son DePerroAI UX Design Lead,Capgemini InventEfi RailiSafety Authority,Technology and Innovation,Capgemini EngineeringAcknowledgementsAnimashree(Anima)AnandkumarBren Professor of Computing and Mathematical Sciences,California Institute of Technology(Caltech)Nebahat ArslanDirector,Group General Couns

178、el and Partnership Officer,Women in AIRicardo Baeza-YatesDirector of Research,Institute for Experiential AI,Northeastern UniversityAustin BaikResponsible AI/AI Governance,TikTok Amir BanifatemiCo-Founder and Director,AI CommonsWilliam BartholomewDirector of Public Policy,Responsible AI,MicrosoftPete

179、 BernardExecutive Director,tinyML FoundationStella BidermanExecutive Director,EleutherAIFrancis BilodeauAssociate Deputy Minister,Innovation,Science and Economic Development CanadaDavor BonaciExecutive Vice-President and Chief Technology Officer,DatastaxMatt BoulosHead,Policy and Safety,ImbueFabio C

180、asatiLead,AI Trust and Governance Lab,ServiceNowJennifer ChayesDean of the College of Computing,Data Science,and Society,University of California,BerkeleyKevin ChungChief Operating Officer,Writer AIJeff CluneAssociate Professor,Department of Computer Science,Faculty of Science,Vector InstituteCathy

181、CobeyGlobal Responsible AI Co-Lead,EYClaudionor CoelhoChief AI Officer,ZscalerSakyasingha DasguptaFounder and Chief Executive Officer,EdgeCortixUmeshwar DayalSenior Fellow and Senior Vice-President,Hitachi America;Corporate Chief Scientist,HitachiMona DiabDirector of Language Technologies Institute,

182、Carnegie Mellon UniversityNavigating the AI Frontier:A Primer on the Evolution and Impact of AI Agents23Yawen Duan Technical Program Manager,Concordia AIMennatallah El-AssadyProfessor,ETH ZurichGilles FayadAdvisor,Institute of Electrical and Electronics Engineers(IEEE)Erica FinkleAI Policy Director,

183、MetaLan GuanGlobal Data and AI Lead,Senior Managing Director,AccentureTom GruberFounder,Humanistic AIMarvin GumprechtAI Organizational Design Expert,Volkswagen GroupGillian HadfieldProfessor,School of Government and Policy;Research Professor,Computer Science,Whiting School of Engineering,Johns Hopki

184、ns UniversityPeter HallinanLeader,Responsible AI,Amazon Web ServicesOr HiltchChief Data and AI Architect,JLLMarius HobbhahnDirector and Co-Founder,Apollo ResearchBabak HodjatChief Technology Officer AI,CognizantSara HookerVice-President,Research,CohereJenia JitsevScientific Lead;Founder,LAIONDavid K

185、anterFounder and Executive Director,MLCommonsVijay KarunamurthyHead of Engineering/Vice-President Engineering,Scale AISean KaskChief AI Strategy Officer,SAPRobert KatzVice-President,Responsible AI and Tech,SalesforceMichael KearnsFounding Director,Warren Center for Network and Data Sciences,Universi

186、ty of PennsylvaniaSteven KellyChief Trust Officer,Institute for Security and TechnologyPrince KohliChief Technology Officer,Automation AnywhereJin KuChief Technology Officer,SendbirdSophie LebrechtChief of Operations and Strategy,Allen Institute for AIAiden LeeCo-Founder and Chief Technology Officer

187、,Twelve LabsStefan LeichenauerVice-President,Engineering,SandboxAQTze Yun LeongProfessor of Computer Science,National University of Singapore Scott LikensGlobal AI and Innovation Technology Lead,PricewaterhouseCoopersShane LukeVice-President of Product and Engineering,WorkdayRichard MallahPrincipal

188、AI Safety Strategist,Future of Life InstitutePilar ManchnSenior Director,Engineering,GoogleDarko MatovskiFounder and Chief Executive Officer,causaLensMao MatsumotoHead of NEC Fellow Office,NECMichael MayHead of Data Analytics and Artificial Intelligence,SiemensStefan MeskenVice-President Research,De

189、epLRisto MiikkulainenProfessor of Computer Science,University of Texas at AustinSatwik MishraExecutive Director,Centre for Trustworthy Technology(CTT)Lama NachmanIntel Fellow,Director of Human and AI Systems Research LabNavigating the AI Frontier:A Primer on the Evolution and Impact of AI Agents24Ni

190、shaLead Consultant Responsible AI Office,InfosysMark NitzbergExecutive Director,Center for Human-Compatible AI,University of California,BerkeleyHenrik OhlssonVice-President;Chief Data Scientist,C3.aiVijoy PandeySenior Vice-President,Outshift by CiscoMaria PocoviSenior Director,Research and Developme

191、nt Emotion AI,UniphoreNaveen RaoVice-President,Generative AI,DatabricksVictor RiparbelliCo-Founder and Chief Executive Officer,SynthesiaJason RugerChief Information Security Officer,LenovoDaniela RusDirector,Computer Science and Artificial Intelligence Laboratory,Massachusetts Institute of Technolog

192、y(MIT)Jun SeitaTeam Leader(Principal Investigator),Medical Data Deep Learning Team,RIKENPaul ShawGroup Security Officer,DentsuNorihiro SuzukiChairman of the Board,Hitachi Research Institute,HitachiFabian TheisScience Director,Helmholtz AssociationLi TieyanChief AI Security Scientist,Huawei Technolog

193、iesAnna TumadttirChief Executive Officer,Creative CommonsChris Van PeltCo-Founder and Chief Information Security Officer,Weights&BiasesKush VarshneyIBM Fellow,IBMAnthony VetroPresident,Chief Executive Officer,IEEE Fellow,Mitsubishi Electric Research LaboratoriesLauren WoodmanChief Executive Officer,

194、DataKindYuan XiaohuiSenior Expert,Tencent HoldingsGrace YeeDirector,Ethical Innovation,AI Ethics,AdobeMichael YoungVice-President,Products,Private AILeonid ZhukovVice-President of Data Science,BCGX;Director of BCG Global AI Institute,Boston Consulting Group(BCG)World Economic Forum Hannah RosenfeldS

195、pecialist,Artificial Intelligence and Machine LearningStephanie SmittkampCoordinator,Artificial Intelligence and DataKarla Yee AmezagaLead,Data Policy and AINavigating the AI Frontier:A Primer on the Evolution and Impact of AI Agents25Endnotes1 Vaswani,A.,et al.(2023).Attention is all you need.https

196、:/arxiv.org/abs/1706.037622 World Economic Forum.(2024).AI value alignment:Guiding artificial intelligence towards shared human goals.https:/www.weforum.org/publications/ai-value-alignment-guiding-artificial-intelligence-towards-shared-human-goals/3 Russell,S.,&Norvig,P.(2021)Artificial intelligence

197、:A modern approach(4th ed.).Prentice Hall.4 Anglen,J.(n.d.).What are AI agents?Artificial Intelligence agents capabilities.Rapid Innovation.Retrieved November 1,2024,from https:/www.rapidinnovation.io/post/what-are-ai-agents-agents-in-artificial-intelligence-explained5 International Organization for

198、 Standardization/International Electrotechnical Commission.(2022).ISO/IEC 22989:2023:Information technology artificial intelligence artificial intelligence concepts and terminology.https:/www.iso.org/standard/74296.html6 Russell,S.,&Norvig,P.(2021).Artificial intelligence:A modern approach(4th ed.).

199、Prentice Hall.7 Vaswani,A.,et al.(2023).Attention is all you need.https:/arxiv.org/abs/1706.037628 Sodhani,S.,et al.(2022).An introduction to lifelong supervised learning.https:/arxiv.org/abs/2207.043549 Buffet,O.,Pietquin,O.,&Weng,P.(2020).Reinforcement learning.https:/arxiv.org/pdf/2005.1441910 Ka

200、ufmann,T.,Wenig,P.,Bengs,V.,&Hllermeier,E.(2024).A survey of reinforcement learning from human feedback.https:/arxiv.org/abs/2312.1492511 Neyshabur,B.,Sedghi,H.,&Zhang,C.(2021).What is being transferred in transfer learning?https:/arxiv.org/abs/2008.1168712 Farahani,A.,Pourshojae,B.,Rasheed,K.,&Arab

201、nia,H.(2021).A concise review of transfer learning.https:/arxiv.org/abs/2104.0214413 Russell,S.,&Norvig,P.(2021)Artificial intelligence:A modern approach(4th ed.).Prentice Hall.14 Ibid.15 Ibid.16 Ibid.17 World Economic Forum.(2023).Data equity:Foundational concepts for generative AI.https:/www.wefor

202、um.org/publications/data-equity-foundational-concepts-for-generative-ai/18 Zhao,P.,Jin,Z.,&Cheng,N.(2023).An in-depth survey of large language model-based artificial intelligence agents.https:/arxiv.org/abs/2309.1436519 Wei,J.,et al.(2023).Chain-of-thought prompting elicits reasoning in large langua

203、ge models.https:/arxiv.org/abs/2201.11903 20 Simankov,V.,Onishchenko,S.,Buchatskiy,P.,&Teploukhov,S.(2023).An approach to the definition of system intelligence in the management of complex systems.IEEE Xplore.https:/ieeexplore.ieee.org/document/10159122 21 Wang,J.,et al.(2024).Mixture-of-agents enha

204、nces large language model capabilities.https:/arxiv.org/abs/2406.04692 22 Pendleton,S.D.et al.(2017).Perception,planning,control,and coordination for autonomous vehicles.Machines,5(1).MDPI.https:/ Mahela,O.,et al.(2022).Comprehensive overview of multi-agent systems for controlling smart grids.CSEE J

205、ournal of power and energy systems,8(1).SciOpen.24 Ibid.25 Ibid.26 LangGraph.(n.d.).Conceptual guide for multi-agent systems.Retrieved November 1,2024,from https:/langchain-ai.github.io/langgraph/concepts/multi_agent/27 US Department of Transportation(2019).Vehicle-to-everything(V2X)communications.h

206、ttps:/www.transportation.gov/v2x 28 Han,S.,et al.(2024).LLM multi-agent systems:Challenges and open problems.https:/arxiv.org/abs/2402.03578 29 Saha,H.,Venkataraman,V.,Speranzon,A.,&Sarkar,S.(2019).A perspective on multi-agent communication for information fusion.https:/arxiv.org/abs/1911.03743v1 30

207、 Zhu,C.,Dastani,M.,&Wang,S.(2024).A survey of multi-agent reinforcement learning with communication.https:/arxiv.org/abs/2203.08975 31 Baroni,M.,Dessi,R.,&Lazaridou,A.(2022).Emergent language-based coordination in deep multi-agent systems.ACL Anthology.https:/aclanthology.org/2022.emnlp-tutorials.3/

208、Navigating the AI Frontier:A Primer on the Evolution and Impact of AI Agents2632 Agentic AI Safety Experts Focus Group.(2024).Guidelines for agentic AI safety:Volume 1.https:/e-space.mmu.ac.uk/635454/1/Safer%2BAgentic%2BAI%2BFoundations%2BPublicationDraft%2BI1D1Jly24.docx%2B%5BIssue%2B2%5D%20%281%29

209、.pdf 33 Royal Academy of Engineering,National Engineering Policy Centre.(2023).Towards autonomous systems in healthcare.https:/nepc.raeng.org.uk/media/mmfbmnp0/towards-autonomous-systems-in-healthcare_-jul-2023-update.pdf 34 World Economic Forum.(2022).Earning digital trust:Decision-making for trust

210、worthy technologies.https:/www3.weforum.org/docs/WEF_Earning_Digital_Trust_2022.pdf 35 Agentic AI Safety Experts Focus Group.(2024).Guidelines for agentic AI safety:Volume 1.https:/e-space.mmu.ac.uk/635454/1/Safer%2BAgentic%2BAI%2BFoundations%2BPublicationDraft%2BI1D1Jly24.docx%2B%5BIssue%2B2%5D%20%

211、281%29.pdf36 Chan,A.,et al.(2024).Visibility into AI agents.https:/arxiv.org/abs/2401.13138 37 Gabriel,I.,et al.(2024).The ethics of advanced AI assistants.https:/arxiv.org/abs/2404.16244 38 Ibid.39 Ibid.40 Ibid.41 AI Seoul Summit(2024).International scientific report on the safety of advanced AI:In

212、terim report.https:/assets.publishing.service.gov.uk/media/6716673b96def6d27a4c9b24/international_scientific_report_on_the_safety_of_advanced_ai_interim_report.pdf 42 Chan,A.,et al.(2024),Visibility into AI agents.https:/arxiv.org/abs/2401.1313843 Royal Academy of Engineering/National Engineering Po

213、licy Centre.(2023).Autonomous systems:A workshop on cross-cutting governance.https:/nepc.raeng.org.uk/media/2hsh552k/as-workshop-report-v4.pdf 44 Ibid.45 Chan,A.,et al.(2024),Visibility into AI agents.https:/arxiv.org/abs/2401.1313846 AI Seoul Summit(2024).International scientific report on the safe

214、ty of advanced AI:Interim report.https:/assets.publishing.service.gov.uk/media/6716673b96def6d27a4c9b24/international_scientific_report_on_the_safety_of_advanced_ai_interim_report.pdf47 World Economic Forum.(2024).Advancing data equity:An action-oriented framework.https:/www.weforum.org/publications

215、/advancing-data-equity-an-action-oriented-framework/48 Risk is defined as the combination of the probability of an occurrence of harm and the severity of that harm if it occurs.See:International Organization for Standardization/International Electrotechnical Commission.(2022).ISO/IEC 22989:2023:Info

216、rmation technology artificial intelligence artificial intelligence concepts and terminology.https:/www.iso.org/standard/74296.html49 Chan,A.,et al.(2024),Visibility into AI agents.https:/arxiv.org/abs/2401.1313850 Royal Academy of Engineering/National Engineering Policy Centre(2020).Safety and ethic

217、s of autonomous systems:Project overview.https:/nepc.raeng.org.uk/media/nqnhktgq/nepc-safety-and-ethics-of-autonomous-systems.pdf 51 World Economic Forum.(2024).AI value alignment:Guiding artificial intelligence towards shared human goals.https:/www.weforum.org/publications/ai-value-alignment-guidin

218、g-artificial-intelligence-towards-shared-human-goals/52 Doshi-Velez,F.,&Kim,B.(2017).Towards a rigorous science of interpretable machine learning.https:/arxiv.org/abs/1702.08608.53 Chan,A.,et al.(2024).Visibility into AI agents.https:/arxiv.org/abs/2401.1313854 Royal Academy of Engineering,National

219、Engineering Policy Centre.(2023).Towards autonomous systems in healthcare.https:/nepc.raeng.org.uk/media/mmfbmnp0/towards-autonomous-systems-in-healthcare_-jul-2023-update.pdf55 World Economic Forum.(n.d.).Safe systems and technologies.AI Governance Alliance.Retrieved October 4,2024,from https:/init

220、iatives.weforum.org/ai-governance-alliance/safesystemsNavigating the AI Frontier:A Primer on the Evolution and Impact of AI Agents27World Economic Forum9193 route de la CapiteCH-1223 Cologny/GenevaSwitzerland Tel.:+41(0)22 869 1212Fax:+41(0)22 786 2744contactweforum.orgwww.weforum.orgThe World Economic Forum,committed to improving the state of the world,is the International Organization for Public-Private Cooperation.The Forum engages the foremost political,business and other leaders of society to shape global,regional and industry agendas.