In collaboration with

Capgemini

Navigating the AI Frontier:

A Primer on the Evolution and
Impact of AI Agents
WHITE PAPER
DECEMBER 2024
Images: Getty Images

Contents
Foreword 3

Executive summary 4

Introduction 5

1 Definition of an AI agent 6

2 The evolution of AI agents 8

2.1. Key technological trends 9

2.2. Types of AI agents 10

2.3. Advanced AI agents 12

2.4. AI agent system 13

2.5. The future of AI agents: Towards multi-agent systems 14

3 Looking ahead 17

3.1. Key benefits 18

3.2. Examples of risks and challenges 18

3.3. Addressing the risk and challenges 20

Conclusion 22

Contributors 23

Endnotes 26

Disclaimer
This document is published by the
World Economic Forum as a contribution
to a project, insight area or interaction.
The findings, interpretations and
conclusions expressed herein are a result
of a collaborative process facilitated and
endorsed by the World Economic Forum
but whose results do not necessarily
represent the views of the World Economic
Forum, nor the entirety of its Members,
Partners or other stakeholders.
© 2024 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.

Navigating the AI Frontier: A Primer on the Evolution and Impact of AI Agents 2

December 2024 Navigating the AI Frontier: A Primer on the
Evolution and Impact of AI Agents

Foreword
Fernando Alvarez
Jeremy Jurgens
Chief Strategy and
Managing Director, World
Development Officer,
Economic Forum
Capgemini

In the contemporary world, where technology is attention and proactive management. Ensuring
rapidly reshaping every aspect of our lives, that AI development aligns with societal values
AI agents are emerging as transformative tools and aspirations is paramount to its successful
that are redefining human interactions and the integration into daily life. The aim of these
operation of our society. These agents, which innovations is to amplify human ingenuity – not to
began as simple computer programs, have replace it – within our economy.
evolved into sophisticated systems with the
capability for autonomous decision-making. This comprehensive overview serves as an
This evolution signifies a major shift, positioning important resource for those involved in shaping the
AI agents as active participants in crucial sectors future of AI technology. By exploring the capabilities
such as healthcare, education, financial services and implications of AI agents, stakeholders can
and beyond. better understand how to leverage the power of
these systems to drive meaningful progress across
The advancement of AI agents brings with it a various sectors. It is through this understanding
wealth of exciting possibilities and transformative that we can ensure AI technologies are developed
potential. Their ability to manage complex tasks responsibly and used in ways that enhance human
with minimal human intervention offers the promise well-being. With careful stewardship, AI agents can
of significantly increased efficiency and productivity. become invaluable allies in fostering innovation and
improving quality of life worldwide.
However, as we step into this AI-driven era, it
is essential to not only harness the immense In partnership, the World Economic Forum and
benefits these technologies offer, but also to Capgemini have joined forces through the AI
address the challenges they present. Issues Governance Alliance to advance this critical topic in
such as ethical considerations require careful collaboration with the AI community.

Navigating the AI Frontier: A Primer on the Evolution and Impact of AI Agents 3

Executive summary
This 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 sectors such as healthcare, customer service
upon their environment to achieve goals, artificial and education. However, AI agents also present
intelligence (AI) agents are being deployed in a wide novel risks, including potential misalignment,2
range of roles in different industries. This requires along with ethical concerns about transparency
the adaptation of governance frameworks to ensure and accountability.
responsible adoption.
Future advances in the area are likely to involve
AI agents, comprising components such as multi-agent systems (MAS), where AI agents
sensors and effectors, have evolved from rule- collaborate to address complex challenges such as
based systems to advanced models capable urban traffic management. More advanced systems
of complex decision-making and independent introduce new demands for interoperability and
operation. Enabled by breakthroughs in deep communication standards to function effectively,
learning, reinforcement learning and the transformer while these protocols still need to be debated and
architecture,1 AI agents span applications from agreed upon by a wider community.
workflow automation to personal assistants. This
progression now encompasses more sophisticated This paper highlights the need for robust
utility-based AI agents that incorporate memory, governance, ethical guidelines and a cross-sectoral
planning and tool integration, broadening their consensus to integrate AI agents safely into society.
capabilities and relevance. As more advanced AI agents continue to proliferate,
it is imperative that their transformative potential
The benefits of AI agents include productivity gains, remains balanced with essential safety, security and
specialized support and improved efficiency in governance considerations.

Navigating the AI Frontier: A Primer on the Evolution and Impact of AI Agents 4

Introduction
AI agents are becoming more advanced, with
significant implications for decision-making,
accountability and oversight.

As artificial intelligence (AI) continues to advance As AI agents continue to advance, society is

and integrate into various sectors of the economy gradually progressing towards the development
and society, understanding the role of AI agents, of innovative systems with increased autonomy,
their capabilities and likely impact is crucial capable of completing tasks with minimal human
for business leaders, policy-makers and other involvement or guidance. This heralds a new era of
stakeholders involved in shaping the future of AI AI-driven innovation and efficiency with the potential
development, implementation and governance. to affect every sector of the global economy. Given
this far-reaching prospect, it is crucial to consider
The concept of an agent – an entity that perceives safety and governance measures to guide the
its environment through sensors and acts on it responsible development and implementation of
through effectors – has been constantly evolving advanced AI agents.4
since the beginning of task automation.3 With
recent advances in large language models (LLM – This paper first defines the concept of AI agents
AI models that process natural language) and large before outlining different types of agents and their
multimodal models (LMM – AI models that process evolution over time. The last section looks ahead
natural language, images, video and/or audio), the and summarizes examples of emerging technical
concept of AI agents is moving into a new phase of and the socioeconomic implications of deploying
rapid development and experimentation. This phase AI agents – along with possible measures to
is currently seeing the emergence of a range of mitigate risks.
novel use cases from coding assistants to workflow
automation, personal assistants and many more
areas of application.

Navigating the AI Frontier: A Primer on the Evolution and Impact of AI Agents 5

1 Definition of an AI agent
An AI agent responds autonomously to
inputs and its reading of its environment
to make complex decisions and change
the environment.

Navigating the AI Frontier: A Primer on the Evolution and Impact of AI Agents 6

 Based on the definition of the International independently and make decisions without
Organization for Standardization,5 an AI agent constant human intervention) and authority
can be broadly defined as an entity that senses (defined as the granted permissions and
percepts (sound, text, image, pressure etc.) access rights to perform specific actions within
using sensors and responds (using effectors) defined boundaries) to take actions to achieve
to its environment. AI agents generally have a set of specified goals, thereby modifying
the autonomy (defined as the ability to operate their environment.

FIGURE 1: The core components of an AI agent

AI agent

Percepts Sensors

Environment

User input Control centre

Digital Physical
infrastructure infrastructure

Effectors

Actions

Source: World Economic Forum

Figure 1 highlights how an agent is made up of information, make decisions and plan actions.
several core components, including: Based on the capabilities of the AI agent, the
control centre involves complex algorithms and
– User input: the external (e.g. human, another models that allow the agent to evaluate different
agent) input that the AI agent receives. This options and choose the best course of action.
could be instructions such as typing via a chat-
based interface, voice-based commands or – Percepts: the data inputs that the AI agent
pre-recorded data. receives about its environment, which could
come from various sensors or other data
– Environment: the bounds in which the AI sources. They represent the agents’ perception
agent operates. It serves as the area in which or understanding of its environment.
the agent applies its sensors and effectors to
percept and modify its surroundings based on – Effectors: the tools an agent uses to take
the inputs received and the actions decided actions upon its environment. In physical
upon by the control centre. The environment environments, effectors might include robotic
can be physical infrastructure such as the arms or wheels, while in the digital environment,
mapped area of an autonomous vehicle or they could be commands sent to other software
digital infrastructure such as the intranet of a systems, such as generating a data visualization
business for a coding agent. or executing a workflow.

– Sensors: mechanisms through which the – Actions: represent the alterations made by
agent perceives its environment. Sensors can effectors. In physical environments, actions
range from physical devices (e.g. cameras or might be pushing an object, whereas in digital
microphones) to digital ones (e.g. queries to environments they could be linked to updating
databases or web services). a database.

– Control centre: typically makes up the core

of the AI agent along with the model, such
as an LLM. The control centre helps process

Navigating the AI Frontier: A Primer on the Evolution and Impact of AI Agents 7

2 The evolution of
AI agents
Developers have transformed AI from rule-based
systems to active agents capable of learning
and adapting while engaged in a task.

Navigating the AI Frontier: A Primer on the Evolution and Impact of AI Agents 8

 The development of AI agents began in the more widespread. AI systems began to learn
1950s,6 and since then they have evolved from from data, adapt over time and improve
simple rule-based systems to sophisticated performance. The introduction of neural
autonomous entities capable of complex networks during this period laid the foundation
decision-making. Early AI was characterized by for deep learning, which has since become
deterministic behaviour, relying on fixed rules and essential to modern AI.
logic that made these systems predictable but
unable to learn or adapt from new experiences. Since 2017, the rise of LLMs has transformed
AI’s capabilities in natural language understanding
Advances in AI research introduced systems and generation. These models use vast amounts
that could handle larger datasets and manage of data to produce human-like text and engage in
uncertainty, leading to probabilistic outcomes and complex language-based tasks.
non-deterministic behaviour. This shift enabled
more flexible and dynamic decision-making, Today’s AI agents use various learning
moving beyond rigid frameworks. techniques, including reinforcement learning, or
transfer learning, allowing them to continuously
The 1990s marked a significant turning point, as refine their abilities, adapt to new environments
machine learning applications became and make more informed decisions.

2.1 Key technological trends

Over the past 25 years, the increase in computing 1. Supervised learning: facilitates learning from
capacity, the availability of large quantities labelled datasets, so the model can accurately
of data on the internet and novel algorithmic predict or classify new, previously unseen data.8
breakthroughs have enabled significant
developments in the base technologies behind 2. Reinforcement learning: enables agents
recent advances in the capabilities of AI agents. to learn optimal behaviours through trial and
These are briefly described below. error in dynamic environments. Agents can
continuously update their knowledge base
without needing periodic retraining.9
Large models
3. Reinforcement learning with human
feedback: enables agents to adapt and
Large language models (LLM) and large multimodal improve through human feedback, specifically
models (LMM) have revolutionized the capabilities focusing on aligning AI behaviour with human
of AI agents, particularly in natural language values and preferences.10
processing and the generation of text, image, audio
and video. 4. Transfer learning: involves taking a pretrained
model, typically trained on a large dataset (e.g.
The emergence of large models has been driven to recognize cars) and adapting it to a new but
by several technological advances and by the related problem (e.g. to recognize trucks).11
transformer architecture, which has paved the way
for a deeper understanding of context and word 5. Fine-tuning: involves taking a pretrained model
relationships, considerably improving the efficiency and further training it on a smaller, task-specific
and performance of natural language processing dataset. This process allows the model to retain
tasks.7 In summary, advanced AI models have its foundational knowledge while improving its
enabled better understanding, generation and performance on specialized tasks.12
engagement with natural language.
These and other learning paradigms are often used
in combination and have dramatically expanded the
Machine learning and deep problem-solving capabilities of AI agents in various
areas of application. The evolution of AI agents
learning techniques is detailed in Figure 2, while the agent types are
further expanded in the following section.
A range of techniques have greatly improved AI
models through increased efficiency and greater
specialization. Some examples of machine- and
deep-learning techniques include:

Navigating the AI Frontier: A Primer on the Evolution and Impact of AI Agents 9

 FIGURE 2: Evolution of AI agents’ capabilities

Deterministic Non-deterministic Current state

Technology trend Multi-agent systems

Large models

Machine learning and deep learning techniques

Agent type
Simple reflex Model-based Goal-based Utility-based Future type

Agent examples
Basic anti-virus Smart Advanced Autonomous Smart city
software thermostat chess AI driving traffic planner

Key characteristics Condition– Internal model of Transfer learning Evaluating scenarios Collaborative
action rules the environment and reinforcement to choose the best methodologies that
learning outcome represent the current
state of the art
Source: World Economic Forum

2.2 Types of AI agents

This section outlines different types of AI agent development. AI agents can be considered as either
and traces their evolution, highlighting the key deterministic or non-deterministic, based on their
technological advances that have supported their defining characteristics, which are outlined below.

TA B L E 1 : Defining characteristics of deterministic and non-deterministic AI agents

Deterministic AI agents Non-deterministic AI agents

Rule-based: operate with fixed rules and logic, Data-driven and probabilistic: make
meaning the same input will always produce the decisions based on statistical patterns in
same output. data, with outcomes that are not fixed but
instead are probabilistic.

Predictable behaviour: the decision-making Flexible and adaptive: able to learn from data,
process is transparent and consistent, which adapt to new situations and handle uncertainty,
makes the outcomes predictable. often resulting in varied outcomes for similar inputs.

Limited adaptability: these systems cannot learn Complex decision-making: use algorithms that
from new data or adjust to changes; they follow factor in probabilities, randomness or other non-
only predefined paths. deterministic elements, allowing for more nuanced
and complex behaviours.

Navigating the AI Frontier: A Primer on the Evolution and Impact of AI Agents 10

Type Definition Examples

Simple reflex agents operate based on a perception – Basic spam filters using keyword matching
of their environment, without consideration of past
– Simple chatbots with predefined responses
experiences.13 Instead, they follow predefined rules
to map specific inputs to specific actions. The – Automated email responders that send
implementation of condition–action rules allows for prewritten replies following specific triggers
rapid responses to environmental stimuli.

These early agents are simple rule-based machines

Simple reflex agents or algorithms designed to provide static information
and unable to adapt or change course.

Model-based reflex agents are designed to track – Smart thermostats that optimize energy
parts of their environment that are not immediately usage by adjusting to current and historical
visible to them.14 They do this by using stored temperature data, as well as user preferences
information from previous observations, allowing
– Smart robotic vacuum cleaners that use
them to make decisions based on both current
sensors and maps to navigate efficiently,
inputs and past experiences. By basing their
avoiding obstacles and optimizing cleaning
actions on both current perceptions and their
paths
internal model, these agents are more adaptable
Model-based
reflex agents than simple reflex agents even though they are also – Modern irrigation systems that use sensors to
governed by condition–action rules. collect real-time data on environmental factors
such as soil, moisture, temperature and
precipitation, to optimize water dispensation

Goal-based agents are able to take future – Advanced chess AI engines that have the
scenarios into account. This type of agent goal of winning the game, planning moves
considers the desirability of actions’ outcomes that maximize the probability of success and
and plans to achieve specific goals.15 The considering a long-term strategy
integration of goal-oriented planning algorithms
– Route optimization systems for logistics that
allows the agent to make decisions based on
set goals for efficient delivery and plan optimal
future outcomes, making them suitable for
routes by setting clear priorities
complex decision-making tasks.
Goal-based agents
– Customer service chatbots that set goals
to resolve customer issues and plan
conversation flows to achieve their
goals efficiently

Utility-based agents employ search and planning – Autonomous driving systems that optimize
algorithms to tackle intricate tasks that lack a safety, efficiency and comfort while evaluating
straightforward outcome, thereby going beyond trade-offs such as speed, fuel efficiency and
simple goal achievement. passenger comfort

They use utility functions to assign a weighted score – Portfolio management systems such as robot-
to each potential state, facilitating optimal decision- advisers that make financial decisions based
making in scenarios with conflicting goals or on utility functions that weigh risk, return and
Utility-based agents uncertainty. Rooted in decision theory, this method client preferences
allows for more advanced decision-making in – Healthcare diagnosis assistants that analyse
complex environments. These agents can balance patient medical records, label patient data
multiple, possibly conflicting objectives according to (e.g. tumour detection) and optimize
their relative significance.16 treatment strategy recommendations in
cooperation with doctors

Navigating the AI Frontier: A Primer on the Evolution and Impact of AI Agents 11

 2.3 Advanced AI agents

The architecture of many current AI agents is often overview of the key components leading to current
based on or linked to LLMs, which are configured breakthroughs in AI agents and their growing range
in complex ways. Figure 3 presents a simplified of capabilities.

FIGURE 3: Key components of advanced AI agents

AI agent

Percepts Sensors Learning

Environment

User input Effectors Control centre

Model

Digital Physical Decision- Memory Tools

infrastructure infrastructure making and management
planning

Actions

Source: World Economic Forum

The AI agent begins with user input, which is is a technique where an AI agent systematically
directed to the agent’s control centre. The user processes and articulates intermediate steps to
input could be a prompt given to carry out an reach a conclusion, which enhances the agent’s
instruction. The control centre directs the ability to solve complex problems in a transparent
user input to the model, which forms the core manner, as each step of the model’s underlying
algorithmic foundation of the AI agent. This model reasoning is reproduced in natural language.19
could be an LLM or an LMM, depending on the
application’s needs. The model then processes Memory management is vital for the continuity and
the input data from the user’s instructions to relevance of operations. This component ensures
generate the desired result.17 that the AI agent remembers previous interactions
and maintains context. This is essential for tasks
At the core of the architecture is the control centre, that require historical data to inform decisions or for
a crucial component that manages the flow of maintaining conversational context in chatbots.
information and commands throughout the system.
It acts as the orchestration layer, directing inputs Tools enable the AI agent to access and interact
to the model and routing the output to appropriate with multiple functions or modalities. For example,
tools or effectors. In simple terms, this layer in an online setting, an AI agent could have access
orchestrates the flow of information between 1) to external tools such as web searches to gather
user inputs, 2) decision-making and planning, 3) real-time information and scheduling tools to
memory management, 4) access to tools and 5) the manage appointments and send reminders, as well
effectors of the system enabling action in digital or as project management software to track tasks and
physical environments.18 deadlines. In terms of modalities, an AI agent could
use natural language processing tools alongside
The decision-making and planning component image recognition capabilities to perform tasks
of an AI agent uses the model’s outputs to assist that require understanding of text-based as well as
in decision-making and planning of multistep visual-based data sources.
processes. In this segment, advanced features
such as chain-of-thought (CoT) reasoning are Once decisions are made or plans set, the
implemented, which allows the AI agent to effectors component of the AI agent executes
engage in multistep reasoning and planning. CoT the required actions. This could involve interacting

Navigating the AI Frontier: A Primer on the Evolution and Impact of AI Agents 12

 with the physical world (in robotics), executing a domains. For example, in a healthcare AI agent, the
software function or providing recommendations application layer would translate model outputs into
and decisions to human users. diagnostics, treatment recommendations or medical
alerts through an appropriate user interface.
The learning component is intrinsic to the
model and enables the AI agent to improve its In summary, when the varying components of
performance over time as the model gathers more an advanced AI agent come together, they
input, using machine learning and deep learning represent the agent’s ability to model the
techniques as mentioned in section 2.1. environment, maintain memory or knowledge
storage with beliefs and preferences, as well as
The application layer surrounds the control inherent abilities to learn, plan, make decisions,
centre, models and other components, acting perceive (sense), act (interact) and communicate
as the interface between the AI agent and its with the agent’s surroundings.
environment. It interprets the outputs from the
control centre and adapts them to specific tasks or

Example of an advanced AI agent: AI agent infotainment system

1.
An AI agent in a car’s infotainment system acts personalizes entertainment based on user habits,
as a smart assistant, activated through voice recommends nearby stops such as restaurants
commands to manage navigation, entertainment, or fuel stations and proactively provides updates
climate controls and other vehicle settings. such as low fuel alerts or optimal recharging points
It processes live traffic, weather and driver for electric vehicles – all while ensuring the driver
preferences to optimize routes, suggesting remains focused on the road.
alternatives around delays or hazards. The agent

2.4 AI agent system

An AI agent system is an organized structure that – Central orchestration, which coordinates

integrates multiple heterogeneous (e.g. rule- and calls of agents and manages the inputs
goal-based agents) or homogeneous (e.g. goal- and outputs accordingly
based only) AI agents.20 Each agent is typically
specialized, possessing its own capabilities, The AI agent system is designed to ensure
knowledge and decision-making processes, while that each agent contributes to the overall
sharing data to collaboratively achieve the goal of objective, whether it involves managing complex
the system. real-time processes such as autonomous driving,
optimizing industrial processes or coordinating
Several designs are possible, such as: activities; for example, in smart city infrastructure.
By dividing the workload among specialized
– Mixture-of-agents, where each agent is called agents, the system can handle dynamic
sequentially, with agents processing the outputs environments and adapt to changing conditions,
from each previous agent21 ensuring optimal performance.

Example of an AI agent system: Autonomous vehicle AI agent system

2.
A human user gets into an autonomous vehicle handles the vehicle’s core mechanics, such as
(AV). The AV is comprised of an AI agent system braking, accelerating and steering.22 The AI agent
that includes agents for perception, path planning, infotainment system serves as the interface with
localization for finding its specific place on the road the passenger, and handles elements such as
and control to steer and brake. processing voice commands and adjusting routes,
climate, entertainment or other in-car settings
The perception and localization agents are based on user preferences.23
dedicated to continuously mapping the environment
through sensors, the global positioning system All agents work together in a coordinated and
(GPS) and cameras. The planning agent calculates centralized manner to ensure the vehicle reaches
the optimal trajectory by factoring in real-time traffic, its destination safely and efficiently, prioritizing
weather and road conditions. The control agent both passenger comfort and safety.24

Navigating the AI Frontier: A Primer on the Evolution and Impact of AI Agents 13

 2.5 The future of AI agents: Towards
multi-agent systems

Multi-agent systems (MAS) consist of multiple with the MAS’s objectives. For example, when
independent AI agents as well as AI agent systems autonomous vehicles (AVs) park in a tight
that collaborate, compete or negotiate to achieve space, they communicate to avoid collision.
collective tasks and goals.25 These agents can be In this case, the MAS objective to prevent
autonomous entities, such as software programs accidents aligns with each AV’s goal of safe
or robots, each typically specialized with its own navigation, allowing them to coordinate
set of capabilities, knowledge and decision-making effectively and reach consensus.
processes. This allows agents to perform tasks in
parallel, communicate with one another and adapt – Supervised architecture: In this model, a
to changes in complex environments. “supervisor” agent coordinates interactions
among other agents. It is useful when agents’
The architecture of a MAS is determined by goals diverge, and consensus may be
the desired outcomes and the goals of each unattainable. The supervisor can mediate and
participating agent or system. There are several prioritize the MAS’s objectives while considering
architectural types,26 for example: each agent’s unique goals, thereby finding a
compromise. An example could be when a
– Network architecture: In this set-up, all buyer and seller agent cannot reach agreement
agents or systems can communicate with on a transaction, which is then mediated by an
one another to reach a consensus that aligns AI agent supervisor.

FIGURE 4: Examples of MAS architecture

Network architecture Supervised architecture

AI agent
AI agent
superivsor

AI agent
AI agent
system

AI agent
AI agent
system

AI agent AI agent
system system

Source: World Economic Forum

While current efforts largely focus on developing These agents can communicate and interact within
AI agents within closed environments or specific a broader adaptive system, enabling them to handle
software ecosystems, the future is likely to see both specific tasks and complex situations more
multiple agents collaborating in different domains efficiently than a single agent, or even an AI agent
and applications. In MAS, different types of agent system, could on its own.
could work together to tackle increasingly complex
tasks that require multistep processes, integrating In some cases, multi-agent systems address
expertise from various fields to achieve more the limitations of single-agent systems, such as
sophisticated outcomes. scalability issues, lack of resilience in the event of

Navigating the AI Frontier: A Primer on the Evolution and Impact of AI Agents 14

 failure or errors and limited skill sets. By distributing crucial for applications that need to grow or evolve
tasks among multiple agents, MAS could increase over time without extensive re-engineering.
both efficiency and capability.
In many ways, multi-agent systems can be
In theory, multi-agent systems are highly adaptable, considered as a future type of system that could
as agents can be dynamically added or removed, coordinate agent actions among multiple users
allowing the system to respond to changing or organizations through human-comprehensible
environments and requirements. This scalability is language or to-be-determined AI agent protocols.

FIGURE 5: The structure and relationships among the AI agent, AI agent system and multi-agent system

Advanced AI agent AI agent system Multi-agent system

Sensors Learning
AI AI AI agent AI
agent 1 agent 2 system 1 agent 2

Control centre
Orchestration
Effectors
Model
AI Other
Decision- Memory Tools agent 3 AI agent
making management
and planning

AI Other
agent 3 agents

EXA MP LE: E X AM PLE : E X AM PLE :

1. AI agent infotainment system 2. Fully autonomous vehicle 3. Connected smart city coordination

Source: World Economic Forum

Example of a multi-agent system: Smart city traffic management with vehicle-to-

3. everything (V2X) communication

In a smart city, a multi-agent system (MAS) energy usage. For example, if an accident occurs,
manages traffic flow in real time, using vehicle-to- AI agents can reroute traffic, adjust signal timings,
everything (V2X) communication, enabling vehicles notify emergency services and communicate with
to interact with other vehicles, pedestrians and road vehicles and pedestrians to avoid the area, all with
infrastructure.27 Each traffic signal is controlled by minimal human intervention. This system optimizes
an AI agent system that communicates with nearby traffic flow, improves road safety and reduces
signals, public transport systems, emergency energy consumption by dynamically adapting to
services and parking services to check availability. real-time conditions. For instance, if a parking lot
Vehicles, equipped with their own AI agent is full, the system can direct vehicles to available
system, share data such as speed, location and parking further away, even if it conflicts with the
road conditions, allowing for coordinated actions driver’s and the onboard AI agent’s preference
to enhance road safety, traffic efficiency and for proximity.

Navigating the AI Frontier: A Primer on the Evolution and Impact of AI Agents 15

Interoperability of multi-agent and consistent; however, they may not adapt
well to dynamic environments where new
systems communication needs arise.29

One technical challenge in multi-agent systems – Emergent protocols: these allow agents to
is associated with enabling effective learn how to communicate effectively based
communication between different AI agents on their experiences, often using reinforcement
and AI agent systems.28 In some cases, learning techniques. This enables agents
interactions are limited by the boundaries of to adapt their communication strategies
native application environments, restricting the to changing environments and tasks.30
potential of AI agents to narrower and more However, decoding and understanding
specialized subdomains, where control is more emergent communication remains an ongoing
easily retained. research challenge.31

The interoperability of AI agents relies on common A good understanding of the messages

communication protocols, which are the rules and exchanged between AI agents is essential,
standards governing how AI agents exchange otherwise it could affect the overall reliability of
information. These protocols can generally be multi-agent systems. This inconsistency could
categorized in two types: lead to misunderstandings or misaligned actions
when agents collaborate, especially in complex
– Predefined protocols: these are based on environments requiring precise coordination.
established agent communication languages To enhance the transparency of multi-agent
and ontologies. Since they are predefined, interaction, the information exchanged needs to
the communication patterns are predictable be easily accessible and interpretable by humans.

Navigating the AI Frontier: A Primer on the Evolution and Impact of AI Agents 16

3 Looking ahead
AI agents have 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 Agents 17

3.1 Key benefits

By scaffolding capabilities such as reasoning, Key characteristics of greater autonomy increasingly

planning and self-checking on top of LLMs, more allow AI agents to tackle open-ended, real-world
capable AI agents emerge that hold the potential challenges that at one time were beyond them
to dramatically increase users’ productivity and – for example, helping in scientific discovery,
absolve them from certain tasks. This could improving the efficiency of complex systems such
involve completing tasks beyond users’ skill sets, as supply chains or electrical grids, managing
such as specialized coding, or partially or fully rare and unusual scenarios in processes that are
offloading tedious tasks that can be done more too infrequent to justify traditional automation,
cheaply, quickly and at a greater scale than before. or enabling physical robots that can manipulate
Additionally, the application of AI agents can play objects and navigate physical environments.32
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.

Examples of the benefits of applications of AI agents include:

Software development
AI agents can help generate, run and check code and other artefacts needed, allowing
software developers to focus on higher value-added activities.

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 could help alleviate the workload of clinical specialists
by assisting doctors in developing tailored treatment plans.33

Enhanced customer experience

AI agent-based chatbots or virtual assistants can offer personalized, round-the-
clock support, increasing customer satisfaction. They have the potential to provide
consistently accurate responses, helping businesses maintain communication quality
and resolve customer issues efficiently.34

Education
AI agents could help personalize learning experiences by adapting content to each
student’s needs, offering real-time feedback and supporting teachers with grading 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
trends, providing faster and more accurate insights for decision-making.

Navigating the AI Frontier: A Primer on the Evolution and Impact of AI Agents 18

3.2 Examples of risks and challenges

While AI agents have the potential to offer numerous – Goal misgeneralization: When AI agents
benefits, they also come with inherent risks, as apply their learned goals inappropriately to
well as novel safety and security implications. For new or unforeseen situations.39
example, an AI system independently pursuing
misaligned objectives could cause immense – Deceptive alignment: When AI agents
harm, especially in scenarios where the AI agents’ appear to be aligned with the intended goals
level of autonomy increases while the level of during training or testing, but their internal
human oversight decreases. AI agents learning objectives differ from what is intended.40
to deceive human operators, pursuing power-
seeking instrumental goals or colluding with other – Malicious use and security vulnerabilities: AI
misaligned agents in unexpected ways could pose agents can amplify the risk of fraud and scams
entirely novel risks.35 increasing both in volume and sophistication.
More capable AI agents can facilitate the
Agent-specific risks can be both technical and generation of scam content at greater speeds
normative. Challenges associated with AI agents and scale than previously possible, and AI
stem from technical limitations, ethical concerns agents can facilitate the creation of more
and broader societal impacts often associated convincing and personalized scam content.
with a system’s level of autonomy and the overall For example, AI systems could help criminals
potential of its use when humans are removed evade security software by correcting language
from the loop. Without a human in the loop at errors and improving the fluency of messages
appropriate steps, agents may take multiple that might otherwise be caught by spam
consequential actions in rapid succession, which filters.41 More capable AI agents could automate
could have significant consequences before a complex end-to-end tasks that would lower the
person notices what is happening.36 point of entry for engaging in harmful activities.
Some forms of cyberattacks could, for example,
AI agents can also amplify known risks associated be automated, allowing individuals with little
with the domain of AI and could introduce entirely domain knowledge or technical expertise to
new risks that can be broadly categorized into execute large-scale attacks.42
technical, socioeconomic and ethical risks.
– Challenges in validating and testing complex
AI agents: The lack of transparency and non-
Technical risks deterministic behaviour of some AI agents
creates significant challenges for validation
and verification. In safety-critical applications,
Examples of technical risks include: this unpredictability complicates efforts to
assure system safety, as it becomes difficult
– Risks from malfunctions due to AI agent to demonstrate reliable performance in all
failures: AI agents can amplify the risks from scenarios.43 While failures in agent-based
malfunctions by introducing new classes of systems are expected, the varied ways in which
failure modes. LLMs, for example, can enable they can fail adds further complexity to safety
agents to produce highly plausible but incorrect assurance. Failsafe mechanisms are essential
outputs, presenting risks in ways that were but could be harder to design due to uncertainty
not possible with earlier technologies. These on potential failure modes.44
emerging failure modes add to traditional issues
such as inaccurate sensors or effectors and Socioeconomic risks
encompass capability- and goal-related failures,
as well as increased security vulnerabilities that
could lead to malfunctions.37 Examples of socioeconomic risks include:

Capability failures occur when an AI agent fails – Over-reliance and disempowerment:

to perform the tasks it was designed for, due to Increasing autonomy of AI agents could reduce
limitations in its ability to understand, process human oversight and increase the reliance on AI
or execute the required actions. Goal-related agents to carry out complex tasks, even in high-
failures occur when a system is highly capable stakes situations. Malfunctions of the AI agents
but nevertheless pursues the wrong goal. These due to design flaws or adversarial attacks may
issues can be caused by: not be immediately apparent if humans are not
in the loop. Additionally, disabling an agent
– Specification gaming: When AI agents could be difficult if a user lacks the required
exploit loopholes or unintended shortcuts expertise or domain knowledge.45
in their programming to achieve their
objectives, rather than fulfilling their goals.38 Pervasive interaction with intelligent AI agents
could also have long-term impacts on individual

Navigating the AI Frontier: A Primer on the Evolution and Impact of AI Agents 19

 and collective cognitive capabilities. For Ethical risks
example, increased reliance on AI agents for
social interactions, such as virtual assistants, AI
agent companions, therapists and so on could Examples of ethical risks include:
contribute to social isolation and possibly affect
mental well-being over time. – Ethical dilemmas in AI decision-making:
The autonomous nature of AI agents raises
– Societal resistance: Resistance to the ethical questions about their decision-making
employment of AI agents could hamper their capabilities in critical situations.
adoption in some sectors or use cases.
– Challenges in ensuring AI transparency
– Employment implications: The use of AI and explainability: Many AI models operate
agents is likely to transform a variety of jobs by as “black boxes”, making decisions based
automating many tasks, increasing productivity on complex and opaque processes, thereby
and altering the skills required in the workforce, making it difficult for users to understand
thus causing partial job displacement. or interpret how decisions are made.46 A
Such displacement could primarily affect lack of transparency could lead to concerns
sectors reliant on routine and repetitive about potential errors or biases in the AI
tasks, in industries such as manufacturing or agent’s decision-making capabilities, which
administrative services. would hinder trust and raise issues of moral
responsibility and legal accountability for
– Financial implications: Organizations decisions made by the AI agent.
could face higher costs associated with the
deployment of AI agents, such as expenses for
securing software systems against cyberthreats
and managing associated operational risks.

3.3 Addressing the risk and challenges

To enable the autonomy of AI agents for cases Within the context of a specific application
where it would greatly improve outcomes, and environment, it is important to adopt a
several challenges must be addressed. These risk analysis methodology that systematically
challenges include safety and security-related identifies, categorizes and assesses all of the
assurance, regulation, moral responsibility and legal risks associated with the AI agent. Such an
accountability, data equity considerations, data approach helps ensure that appropriate and
governance and interoperability, skills, culture and effective mitigation mechanisms and strategies
perceptions.47 Addressing these challenges requires can be implemented by relevant stakeholders
a comprehensive approach throughout the stages at the technical, socioeconomic and ethical levels.
of design, development, deployment and use of
AI agents as well as changes across policy and
regulation. As advanced AI agents and multi-agent Technical risk measures
systems continue to evolve and integrate
into various aspects of digital infrastructure,
associated governance frameworks that take Examples of technical risk measures:
increasingly complex scenarios into consideration
need to be established. – Improving information transparency: Where,
why, how, and by whom information is used
In assessing and mitigating the risks of potential is critical for understanding how a system
harm from AI agents, it is essential to understand operates and why certain decisions are made
the specific application and environment of the AI by the agent. Measures can be implemented
agent (including stakeholders that may be affected). to improve the transparency of AI agents such
The risks of potential harm from an AI agent stem as the integration of behavioural monitoring
largely from the context in which it is deployed.48 and implementation of thresholds, triggers and
In high-stakes environments such as healthcare or alerts that involve continuous observation and
autonomous driving, even small errors or biases can analysis of the agent’s actions and decisions.
lead to significant consequences for the users of Implementing behavioural monitoring helps to
such systems. Conversely, in low-stakes contexts, ensure that failures are better understood and
such as customer service, the same AI agent might properly mitigated when they occur.49
pose minimal risks, as mistakes are less likely to
cause serious harm.

Navigating the AI Frontier: A Primer on the Evolution and Impact of AI Agents 20

Socioeconomic risk measures Ethical risk measures

Examples of socioeconomic risk measures: Examples of ethical risk measures:

– Public education and awareness: Developing – Clear ethical guidelines: Prioritizing human
and executing strategies to inform and engage rights, privacy and accountability are essential
the public are essential to mitigate the risks of measures to ensure that AI agents make
over-reliance and disempowerment in social decisions that are aligned with human and
interactions with AI agents. These efforts societal values.51
should aim to equip individuals with a solid
understanding of the capabilities and limitations – Behavioural monitoring: Implementing
of AI agents, allowing for more informed measures that allow users to trace and
interactions, along with healthy integrations. understand the underlying reasoning
behind an AI agent’s decisions is necessary
– A forum to collect public concerns: to mitigate transparency challenges.52
Acceptance and involvement, trust and Behavioural monitoring can make system
psychological safety are crucial to tackle behaviour and decisions visible and
societal resistance and for the proper adoption interpretable, which enhances overall user
and integration of AI agents into various understanding of interactions. This approach
processes. Without sufficient human “buy-in”, also strengthens the governance structure
the implementation of AI agents would face surrounding AI agents and helps increase
significant challenges. In addressing societal stakeholder accountability.53
resistance and creating wider trust in AI agents
and autonomous systems, it is important that As the adoption of AI agents increases, critical
public concerns are heard and addressed trade-offs need to be made. Given the complex
throughout the design and deployment of nature of many advanced AI agents, safety should
advanced AI agents.50 be regarded as a critical factor alongside other
considerations such as cost and performance,
– Thoughtful strategies for deployment: intellectual property, accuracy, and transparency,
Organizations can embrace deliberate as well as implied social trade-offs when it comes
strategies around increased efficiency and task to deployment.
augmentation rather than focusing on outright
worker replacement efforts. By prioritizing The level of autonomy of advanced AI agents is
proactive measures such as retraining likely to continue to increase due to ever more
programmes, workers can be supported in capable models and reasoning capabilities.54 The
transitioning to new or changed roles. complexities of more advanced systems call for
a multidisciplinary approach that includes 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
sophisticated 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 AI Frontier: A Primer on the Evolution and Impact of AI Agents 21

Conclusion
AI agents are becoming more autonomous in
their operation and decision-making, bringing
potential benefits and risks.

The development of AI agents has been marked The rapid advance of AI agent capabilities is set to
by significant milestones, from the early days of be followed by a wave of innovation in AI agents,
simple reflex agents to sophisticated multi-agent which could have the ability to transform the global
systems. Recent advances in LLMs and LMMs economy and the roles of human labour in new and
have resulted in the next evolution of AI agents, significant ways.
which have moved from basic systems that react
to immediate stimuli to complex entities capable of Further research is necessary to explore the
planning, learning and making decisions based on a safety, security and societal impacts of AI agents
comprehensive understanding of their environment and multi-agent systems, emphasizing both
and user needs. technical solutions and organizational governance
frameworks. These efforts are critical for mitigating
The ongoing development of AI agents is risks associated with the ongoing development,
fundamentally linked to increased autonomy, deployment and increasing use of more
improved learning capabilities, enhanced decision- sophisticated AI agents in a range of domains.
making abilities and multi-agent collaboration. As
the architecture and emerging use cases for AI At this point, it is vital for stakeholders to come
agents continue to proliferate, the shift towards together throughout technical, civil society,
multi-agent systems that can collaborate in applied and governance-facing communities to
increasingly complex environments is likely research, discuss and build consensus on novel
to continue. governance mechanisms.

Increased autonomy plays an important part This white paper has offered an initial exploration of
in the evolution of AI agents and creates novel the rapidly evolving landscape of AI agents, aiming
opportunities for new applications while also to promote deeper understanding of this emerging
presenting unique risks to society. The introduction field and spark conversation on responsible
of AI agents will likely reduce the need for human adoption and diffusion practices. Through equitable
involvement and oversight in some areas, bringing a development, deployment and governance, the
more efficient approach to tedious tasks. However, a growing presence of advanced AI agents holds the
reduction in human oversight could also increase the promise of driving positive societal transformation
risk of accidents. Furthermore, increased automation for many years to come.
of workflows could be a way for malicious actors to
exploit novel vulnerabilities, while also exacerbating
socioeconomic and ethical risks.

Navigating the AI Frontier: A Primer on the Evolution and Impact of AI Agents 22

Contributors
The World Economic Forum’s AI Governance Alliance Safe Systems and Technologies working group
convenes 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 practices
for AI system design and implementation.55

World Economic Forum Capgemini

Benjamin Larsen Olivier Denti

Lead, Artificial Intelligence and Machine Learning Data Architect, AI, Capgemini Invent

Cathy Li Jason DePerro

Head, AI, Data and Metaverse; Member of the AI UX Design Lead, Capgemini Invent
Executive Committee
Efi Raili
Stephanie Teeuwen Safety Authority, Technology and Innovation,
Specialist, Data Policy and AI Capgemini Engineering

Acknowledgements

Animashree (Anima) Anandkumar Matt Boulos

Bren Professor of Computing and Mathematical Head, Policy and Safety, Imbue
Sciences, California Institute of Technology (Caltech)
Fabio Casati
Nebahat Arslan Lead, AI Trust and Governance Lab, ServiceNow
Director, Group General Counsel and Partnership
Officer, Women in AI Jennifer Chayes
Dean of the College of Computing, Data Science,
Ricardo Baeza-Yates and Society, University of California, Berkeley
Director of Research, Institute for Experiential AI,
Northeastern University Kevin Chung
Chief Operating Officer, Writer AI
Austin Baik
Responsible AI/AI Governance, TikTok Jeff Clune
Associate Professor, Department of Computer
Amir Banifatemi Science, Faculty of Science, Vector Institute
Co-Founder and Director, AI Commons
Cathy Cobey
William Bartholomew Global Responsible AI Co-Lead, EY
Director of Public Policy, Responsible AI, Microsoft
Claudionor Coelho
Pete Bernard Chief AI Officer, Zscaler
Executive Director, tinyML Foundation
Sakyasingha Dasgupta
Stella Biderman Founder and Chief Executive Officer, EdgeCortix
Executive Director, EleutherAI
Umeshwar Dayal
Francis Bilodeau Senior Fellow and Senior Vice-President, Hitachi
Associate Deputy Minister, Innovation, Science and America; Corporate Chief Scientist, Hitachi
Economic Development Canada
Mona Diab
Davor Bonaci Director of Language Technologies Institute,
Executive Vice-President and Chief Technology Carnegie Mellon University
Officer, Datastax

Navigating the AI Frontier: A Primer on the Evolution and Impact of AI Agents 23

Yawen Duan Steven Kelly
Technical Program Manager, Concordia AI Chief Trust Officer, Institute for Security and
Technology
Mennatallah El-Assady
Professor, ETH Zurich Prince Kohli
Chief Technology Officer, Automation Anywhere
Gilles Fayad
Advisor, Institute of Electrical and Electronics Jin Ku
Engineers (IEEE) Chief Technology Officer, Sendbird

Erica Finkle Sophie Lebrecht

AI Policy Director, Meta Chief of Operations and Strategy, Allen Institute for
AI
Lan Guan
Global Data and AI Lead, Senior Managing Director, Aiden Lee
Accenture Co-Founder and Chief Technology Officer, Twelve
Labs
Tom Gruber
Founder, Humanistic AI Stefan Leichenauer
Vice-President, Engineering, SandboxAQ
Marvin Gumprecht
AI Organizational Design Expert, Volkswagen Group Tze Yun Leong
Professor of Computer Science, National University
Gillian Hadfield of Singapore
Professor, School of Government and Policy;
Research Professor, Computer Science, Whiting Scott Likens
School of Engineering, Johns Hopkins University Global AI and Innovation Technology Lead,
PricewaterhouseCoopers
Peter Hallinan
Leader, Responsible AI, Amazon Web Services Shane Luke
Vice-President of Product and Engineering,
Or Hiltch Workday
Chief Data and AI Architect, JLL
Richard Mallah
Marius Hobbhahn Principal AI Safety Strategist, Future of Life Institute
Director and Co-Founder, Apollo Research
Pilar Manchón
Babak Hodjat Senior Director, Engineering, Google
Chief Technology Officer AI, Cognizant
Darko Matovski
Sara Hooker Founder and Chief Executive Officer, causaLens
Vice-President, Research, Cohere
Mao Matsumoto
Jenia Jitsev Head of NEC Fellow Office, NEC
Scientific Lead; Founder, LAION
Michael May
David Kanter Head of Data Analytics and Artificial Intelligence,
Founder and Executive Director, MLCommons Siemens

Vijay Karunamurthy Stefan Mesken

Head of Engineering / Vice-President Engineering, Vice-President Research, DeepL
Scale AI
Risto Miikkulainen
Sean Kask Professor of Computer Science, University of Texas
Chief AI Strategy Officer, SAP at Austin

Robert Katz Satwik Mishra

Vice-President, Responsible AI and Tech, Executive Director, Centre for Trustworthy
Salesforce Technology (CTT)

Michael Kearns Lama Nachman

Founding Director, Warren Center for Network and Intel Fellow, Director of Human and AI Systems
Data Sciences, University of Pennsylvania Research Lab

Navigating the AI Frontier: A Primer on the Evolution and Impact of AI Agents 24

Nisha Anna Tumadóttir
Lead Consultant Responsible AI Office, Infosys Chief Executive Officer, Creative Commons

Mark Nitzberg Chris Van Pelt

Executive Director, Center for Human-Compatible Co-Founder and Chief Information Security Officer,
AI, University of California, Berkeley Weights & Biases

Henrik Ohlsson Kush Varshney

Vice-President; Chief Data Scientist, C3.ai IBM Fellow, IBM

Vijoy Pandey Anthony Vetro

Senior Vice-President, Outshift by Cisco President, Chief Executive Officer, IEEE Fellow,
Mitsubishi Electric Research Laboratories
Maria Pocovi
Senior Director, Research and Development Lauren Woodman
Emotion AI, Uniphore Chief Executive Officer, DataKind

Naveen Rao Yuan Xiaohui

Vice-President, Generative AI, Databricks Senior Expert, Tencent Holdings

Victor Riparbelli Grace Yee

Co-Founder and Chief Executive Officer, Synthesia Director, Ethical Innovation, AI Ethics, Adobe

Jason Ruger Michael Young

Chief Information Security Officer, Lenovo Vice-President, Products, Private AI

Daniela Rus Leonid Zhukov

Director, Computer Science and Artificial Intelligence Vice-President of Data Science, BCGX; Director of
Laboratory, Massachusetts Institute of Technology BCG Global AI Institute, Boston Consulting Group
(MIT) (BCG)

Jun Seita
Team Leader (Principal Investigator), Medical Data World Economic Forum
Deep Learning Team, RIKEN

Paul Shaw Hannah Rosenfeld

Group Security Officer, Dentsu Specialist, Artificial Intelligence and Machine
Learning
Norihiro Suzuki
Chairman of the Board, Hitachi Research Institute, Stephanie Smittkamp
Hitachi Coordinator, Artificial Intelligence and Data

Fabian Theis Karla Yee Amezaga

Science Director, Helmholtz Association Lead, Data Policy and AI

Li Tieyan
Chief AI Security Scientist, Huawei Technologies

Navigating the AI Frontier: A Primer on the Evolution and Impact of AI Agents 25

 Endnotes
1 Vaswani, A., et al. (2023). Attention is all you need. https://arxiv.org/abs/1706.03762
2 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: 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-explained
5 International Organization for Standardization/International Electrotechnical Commission. (2022). ISO/IEC 22989:2023:
Information technology – artificial intelligence – artificial intelligence concepts and terminology. https://www.iso.org/
standard/74296.html
6 Russell, S., & Norvig, P. (2021). Artificial intelligence: A modern approach (4th ed.). Prentice Hall.
7 Vaswani, A., et al. (2023). Attention is all you need. https://arxiv.org/abs/1706.03762
8 Sodhani, S., et al. (2022). An introduction to lifelong supervised learning. https://arxiv.org/abs/2207.04354
9 Buffet, O., Pietquin, O., & Weng, P. (2020). Reinforcement learning. https://arxiv.org/pdf/2005.14419
10 Kaufmann, T., Wenig, P., Bengs, V., & Hüllermeier, E. (2024). A survey of reinforcement learning from human feedback.
https://arxiv.org/abs/2312.14925
11 Neyshabur, B., Sedghi, H., & Zhang, C. (2021). What is being transferred in transfer learning? https://arxiv.org/
abs/2008.11687
12 Farahani, A., Pourshojae, B., Rasheed, K., & Arabnia, H. (2021). A concise review of transfer learning. https://arxiv.org/
abs/2104.02144
13 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.weforum.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/2309.14365
19 Wei, J., et al. (2023). Chain-of-thought prompting elicits reasoning in large language 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 enhances 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://www.mdpi.com/2075-1702/5/1/6#:~:text=Perception%20refers%20to%20the%20ability,by%20the%20
higher%20level%20processes
23 Mahela, O., et al. (2022). Comprehensive overview of multi-agent systems for controlling smart grids. CSEE Journal 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.
https://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 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/

Navigating the AI Frontier: A Primer on the Evolution and Impact of AI Agents 26

32 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.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 trustworthy 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%281%29.pdf
36 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: Interim 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/2401.13138
43 Royal Academy of Engineering/ National Engineering Policy 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/2401.13138
46 AI Seoul Summit (2024). International scientific report on the safety of advanced AI: Interim report. https://assets.
publishing.service.gov.uk/media/6716673b96def6d27a4c9b24/international_scientific_report_on_the_safety_of_
advanced_ai_interim_report.pdf
47 World Economic Forum. (2024). Advancing data equity: An action-oriented framework. https://www.weforum.org/
publications/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: Information technology – artificial intelligence – artificial intelligence concepts and terminology. https://www.
iso.org/standard/74296.html
49 Chan, A., et al. (2024), Visibility into AI agents. https://arxiv.org/abs/2401.13138
50 Royal Academy of Engineering/ National Engineering Policy Centre (2020). Safety and ethics 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-guiding-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/2401.13138
54 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
55 World Economic Forum. (n.d.). Safe systems and technologies. AI Governance Alliance. Retrieved October 4, 2024, from
https://initiatives.weforum.org/ai-governance-alliance/safesystems

Navigating the AI Frontier: A Primer on the Evolution and Impact of AI Agents 27

The 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.

World Economic Forum

91–93 route de la Capite
CH-1223 Cologny/Geneva
Switzerland
Tel.: +41 (0) 22 869 1212
Fax: +41 (0) 22 786 2744
[email protected]
www.weforum.org