Multi-agent AI systems need transparency

As the title indicates, Multiagent Systems: A Modern Approach to Distributed Artificial Intelligence covers the design and development of multiagent and distributed AI systems. The purpose of this book is to provide a comprehensive overview of the field. It is an excellent collection of closely related papers that provides a wonderful introduction to multiagent systems and distributed AI.

This study investigates the implementation of LLM agents in smart city management, leveraging both the inherent language processing abilities of LLMs and the distributed problem solving capabilities of multi-agent systems for the improvement of urban decision making processes. A multi-agent system architecture combines LLMs with existing urban information systems to process complex queries and generate contextually relevant responses for urban planning and management. The research is focused on three main hypotheses testing: (1) LLM agents’ capability for effective routing and processing diverse urban queries, (2) the effectiveness of Retrieval-Augmented Generation (RAG) technology in improving response accuracy when working with local knowledge and regulations, and (3) the impact of integrating LLM agents with existing urban information systems. Our experimental results, based on a comprehensive validation dataset of 150 question–answer pairs, demonstrate significant improvements in decision support capabilities. The multi-agent system achieved pipeline selection accuracy of 94–99% across different models, while the integration of RAG technology improved response accuracy by 17% for strategic development queries and 55% for service accessibility questions. The combined use of document databases and service APIs resulted in the highest performance metrics (G-Eval scores of 0.68–0.74) compared to standalone LLM responses (0.30–0.38). Using St. Petersburg’s Digital Urban Platform as a testbed, we demonstrate the practical applicability of this approach to create integrated city management systems with support complex urban decision making processes. This research contributes to the growing field of AI-enhanced urban management by providing empirical evidence of LLM agents’ effectiveness in processing heterogeneous urban data and supporting strategic planning decisions. Our findings suggest that LLM-based multi-agent systems can significantly enhance the efficiency and accuracy of urban decision making while maintaining high relevance in responses.

This paper presents a multi-agent AI system designed to provide accurate diagnostic and personalized treatment recommendations for heart attack, heart failure, cardiac arrhythmia, coronary artery disease, and left ventricular hypertrophy. The system tackles the challenges of integrating various data sources, including electronic health records (EHR), cardiac imaging, genetic information, and electrocardiogram (ECG) data, within a unified multi-agent framework for personalized care related to these conditions. A collaborative network of specialized AI agents, such as the EHR Agent, Cardiac Imaging Agent, Genetic Analysis Agent, and ECG Analysis Agent, work in concert to process and analyze this multi data, identifying potential cardiac conditions and risk factors associated with the above-mentioned target indicators. Research Questions/Hypothesis: This study investigates whether a multi-agent AI system can effectively process patient data, including symptoms, genetic information, and test results, to generate potential conditions and diagnoses. We hypothesize that this integrated approach can potentially improve the speed of assessment for accurate and timely diagnosis, provide relevant diagnostic information and personalized treatment recommendation. Methods/Approach: The multi-agent system comprises several specialized agents responsible for tasks such as symptom analysis, diagnosis, and treatment planning. The system is targeted at processing patient data, including symptom descriptions and test results from labs (biomarkers), ECG, echo, MRI and CT scans, along with genetic variants. The symptom analysis agent identifies potential cardiovascular conditions based on input symptoms. The diagnostic agent then integrates information from potential conditions, patient history, and test results to generate a diagnosis. Results/Data: Analysis of simulated data demonstrates that the symptom analysis agent consistently identifies expected potential conditions with high level of speed and accuracy. Recording 1-2 seconds of diagnosis time with precision level of 98% based on simulated data and programmed logic. We’re only reporting metrics based on the internal consistency of the agent's logic and simulated outcomes. Conclusion(s): The developed multi-agent system demonstrates a functional approach to integrating diverse simulated patient data for cardiovascular assessment and potential diagnosis.

The article discusses the theoretical foundations of using multi-agent AI systems for sustainable development of the banking sector. The relevance of the study is that in modern conditions there are rapid qualitative changes and large-scale growth in the use of multi-agent systems on artificial intelligence in all areas of activity, including banking, which largely determines the sustainable development of the banking system. The novelty lies in the fact that the study proposes an approach to forming a voice bot — a multi-agent system on artificial intelligence, which acts as a seller, is launched by the client in Telegram by pressing a button, freely conducts a conversation with a potential client by voice, asking and receiving answers to the required questions, forming a sales funnel from the list of "hot clients", ensuring an increase in productivity in the execution of business processes by 40 % or more. The practical significance is that the developed voice bot in the format of a multi-agent system based on artificial intelligence can be successfully implemented in a commercial bank, replacing human labor, ensuring a reduction in operating costs, increasing the efficiency of banking products and services by increasing sales due to higher productivity.

Cutting edge technology in intelligence involves multi agent systems (MAS) which allow autonomous agents to interact in shared environments by either working together or competing to achieve common or individual goals. This study delves into the aspects of cooperation and rivalry in MAS and illustrates their application in practical situations, like autonomous vehicles, robot’s interactions, and financial settings. In addition to that we explore the obstacles like coordination, learning and communication that come up while creating MAS frameworks and how sophisticated algorithms like deep reinforcement learning help in running these agents. By tackling both competitive interactions within MAS our goal is to offer a thorough grasp of the possible uses and upcoming paths, in this area. Emerging technologies like OpenAIs agent models play a significant role in showcasing the changing landscape of MAS and its transformative effects on various industries, like healthcare and defense. Keywords: Multi-Agent Systems (MAS), Autonomous Agents, Collaboration, Competition, Deep Reinforcement Learning, Game Theory, Distributed AI, Swarm Intelligence, Agent-Based Modeling, AI Coordination, Adversarial AI

<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Context</i> : Games are a well-established scenario to test AI and Multi-Agent Systems (MAS) proposals due to their popularity and defiance. However, there is no big picture of the application of this technology to games, the evolution of the kind of problem tackled, or the game scenarios in which agents have been experimented. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Objective</i> : To perform a systematic mapping to characterise the state of the art in the field of MAS applied to virtual games and to identify trends, strengths, and gaps for further research. Method: A Systematic Mapping Study has been conducted to find primary studies in the field. A search was performed on title, abstracts, and keywords, whilst classification, data extraction, and further analysis were performed according to specific criteria focused on MAS papers with experimentation and evidence in a game scenario. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Results</i> : 78 studies published between 1998 and 2021 were found. Studies have been classified according to the MAS problem faced and the agent reasoning strategy. We detect that Machine Learning is the most common AI technique for MAS in games, considering both reinforcement learning and evolutionary techniques. MAS are used in a variety of gaming genres, especially in Real-Time Strategy (RTS), Sports and Simulation. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Conclusions</i> : RTS and Sports games are well-suited for concrete MAS problems such as multi-agent planning and task allocation. Expanding evidence and experimentation on other aspects related to scalability and usability issues is discussed. Those MAS problems and experiments that remain slightly modelled on games or are not thoroughly studied yet have been also identified.

In multi-agent systems, emergent norms and distributed decision-making often produce unanticipated behaviors that complicate traditional AI governance frameworks. This paper introduces an adaptive accountability method that traces responsibility flows among networked agents, continuously detects adverse emergent norms, and intervenes to recalibrate local objectives or policies in near real time. By combining lifecycle-based auditing, decentralized governance, and norm detection algorithms, our approach enables robust oversight in dynamic, evolving environments. To validate its scalability and effectiveness, we conduct a series of large-scale simulation experiments on up to 100 agents using an HPC environment. Our ablation studies—covering multiple seeds, varied penalty settings, and different intervention policies—demonstrate that the framework can preserve high collective reward while significantly reducing inequality. In particular, we show that adaptive interventions prevent harmful collusion or hoarding in over 90% of tested configurations, even under partial observability. These results indicate that our method not only mitigates unforeseen disruptions but also aligns agent behaviors with ethical and legal guidelines at scale. Overall, the resulting framework offers a practical path toward ethically sound, multi-agent AI systems that remain responsive to shifting data distributions, organizational policies, and real-world complexity.

This paper presents a multi-agent AI framework for legal aid, designed to support real-world case fact management through interactive dialogue. Our agent framework emulates the iterative questioning, clarification, and synthesis processes of legal professionals, not solely on the fragmented information initially provided by litigants. By engaging in multi-turn interactions, the system incrementally supplements missing details and mitigates risks of misinterpretation, thereby aligning more closely with the dynamics of real legal consultations. A key contribution of this work is the use of real-world legal aid case records as training and evaluation material, ensuring that the framework is grounded in authentic data rather than synthetic simulations. The system is implemented as a collaboration among specialized agents: (1) litigantTwins, which maintains factual integrity and guards against hallucinations; (2) legalAider, which leverages legal knowledge to generate context-sensitive follow-up questions and update case narratives; and (3) Evaluator, which compares AI-generated case records against ground truth facts, assessing factual correctness through qualitative and quantitative measures. Technically, the framework can be built on LLMs, with multi-agent system and prompt design enabling robust coordination. This architecture enhances the quality of fact construction and reasoning, while also offering a scalable solution for online public legal aid services. Beyond its technical contributions, this research highlights its public value, accessibility, and commitment to fairness in the distribution of legal resources. By integrating multi-agent AI with real-world case data, the framework addresses both the technological and socio-legal dimensions of legal information systems, advancing on legal knowledge management, deployment of conversational agents, and normative reasoning in multi-agent systems.

The main challenge that artificial intelligence research is facing nowadays is how to guarantee the development of responsible technology. And, in particular, how to guarantee that autonomy is responsible. The social fears on the actions taken by AI can only be appeased by providing ethical certification and transparency of systems. However, this is certainly not an easy task. As we very well know in the multiagent systems field, the prediction accuracy of system outcomes has limits as multiagent systems are actually examples of complex systems. And AI will be social, there will be thousands of AI systems interacting among themselves and with a multitude of humans; AI will necessarily be multiagent. Although we cannot provide complete guarantees on outcomes, we must be able to define with accuracy what autonomous behaviour is acceptable (ethical), to provide repair methods for anomalous behaviour and to explain the rationale of AI decisions. Ideally, we should be able to guarantee responsible behaviour of individual AI systems by construction. I understand by an ethical AI system one that is capable of deciding what are the most convenient norms, abide by them and make them evolve and adapt. The area of multiagent systems has developed a number of theoretical and practical tools that properly combined can provide a path to develop such systems, that is, provide means to build ethical-by-construction systems: agreement technologies to decide on acceptable ethical behaviour, normative frameworks to represent and reason on ethics, and electronic institutions to operationalise ethical interactions. Along my career, I have contributed with tools on these three areas. In this keynote, I will describe a methodology to support their combination that incorporates some new ideas from law, and organisational theory.

The main challenge that artificial intelligence research is facing nowadays is how to guarantee the development of responsible technology. And, in particular, how to guarantee that autonomy is responsible. The social fears on the actions taken by AI can only be appeased by providing ethical certification and transparency of systems. (See for instance the Barcelona declaration https://www.iiia.csic.es/barcelonadeclaration/) However, this is certainly not an easy task. As we very well know in the multiagent systems field, the prediction accuracy of system outcomes has limits as multiagent systems are actually examples of complex systems. And AI will be social, there will be thousands of AI systems interacting among themselves and with a multitude of humans; AI will necessarily be multiagent. Although we cannot provide complete guarantees on outcomes, we must be able to define with accuracy what autonomous behaviour is acceptable (ethical), to provide repair methods for anomalous behaviour and to explain the rationale of AI decisions. Ideally, we should be able to guarantee responsible behaviour of individual AI systems by construction. I understand by an ethical AI system one that is capable of deciding what are the most convenient norms, abide by them and make them evolve and adapt. The area of multiagent systems has developed a number of theoretical and practical tools that properly combined can provide a path to develop such systems, that is, provide means to build ethical-by-construction systems: agreement technologies to decide on acceptable ethical behaviour, normative frameworks to represent and reason on ethics, and electronic institutions to operationalise ethical interactions. Along my career, I have contributed with tools on these three areas. In this keynote, I will describe a methodology to support their combination that incorporates some new ideas from law, and organisational theory.

This dialogue is from an early scene in the 2014 film Ex Machina, in which Nathan has invited Caleb to determine whether Nathan has succeeded in creating artificial intelligence.1 The achievement of powerful artificial general intelligence has long held a grip on our imagination not only for its exciting as well as worrisome possibilities, but also for its suggestion of a new, uncharted era for humanity. In opening his 2021 BBC Reith Lectures, titled "Living with Artificial Intelligence," Stuart Russell states that "the eventual emergence of general-purpose artificial intelligence [will be] the biggest event in human history."2Over the last decade, a rapid succession of impressive results has brought wider public attention to the possibilities of powerful artificial intelligence. In machine vision, researchers demonstrated systems that could recognize objects as well as, if not better than, humans in some situations. Then came the games. Complex games of strategy have long been associated with superior intelligence, and so when AI systems beat the best human players at chess, Atari games, Go, shogi, StarCraft, and Dota, the world took notice. It was not just that Als beat humans (although that was astounding when it first happened), but the escalating progression of how they did it: initially by learning from expert human play, then from self-play, then by teaching themselves the principles of the games from the ground up, eventually yielding single systems that could learn, play, and win at several structurally different games, hinting at the possibility of generally intelligent systems.3Speech recognition and natural language processing have also seen rapid and headline-grabbing advances. Most impressive has been the emergence recently of large language models capable of generating human-like outputs. Progress in language is of particular significance given the role language has always played in human notions of intelligence, reasoning, and understanding. While the advances mentioned thus far may seem abstract, those in driverless cars and robots have been more tangible given their embodied and often biomorphic forms. Demonstrations of such embodied systems exhibiting increasingly complex and autonomous behaviors in our physical world have captured public attention.Also in the headlines have been results in various branches of science in which AI and its related techniques have been used as tools to advance research from materials and environmental sciences to high energy physics and astronomy.4 A few highlights, such as the spectacular results on the fifty-year-old protein-folding problem by AlphaFold, suggest the possibility that AI could soon help tackle science's hardest problems, such as in health and the life sciences.5While the headlines tend to feature results and demonstrations of a future to come, AI and its associated technologies are already here and pervade our daily lives more than many realize. Examples include recommendation systems, search, language translators - now covering more than one hundred languages - facial recognition, speech to text (and back), digital assistants, chatbots for customer service, fraud detection, decision support systems, energy management systems, and tools for scientific research, to name a few. In all these examples and others, AI-related techniques have become components of other software and hardware systems as methods for learning from and incorporating messy real-world inputs into inferences, predictions, and, in some cases, actions. As director of the Future of Humanity Institute at the University of Oxford, Nick Bostrom noted back in 2006, "A lot of cutting-edge AI has filtered into general applications, often without being called AI because once something becomes useful enough and common enough it's not labeled AI anymore."6As the scope, use, and usefulness of these systems have grown for individual users, researchers in various fields, companies and other types of organizations, and governments, so too have concerns when the systems have not worked well (such as bias in facial recognition systems), or have been misused (as in deepfakes), or have resulted in harms to some (in predicting crime, for example), or have been associated with accidents (such as fatalities from self-driving cars).7Dædalus last devoted a volume to the topic of artificial intelligence in 1988, with contributions from several of the founders of the field, among others. Much of that issue was concerned with questions of whether research in AI was making progress, of whether AI was at a turning point, and of its foundations, mathematical, technical, and philosophical-with much disagreement. However, in that volume there was also a recognition, or perhaps a rediscovery, of an alternative path toward AI - the connectionist learning approach and the notion of neural nets-and a burgeoning optimism for this approach's potential. Since the 1960s, the learning approach had been relegated to the fringes in favor of the symbolic formalism for representing the world, our knowledge of it, and how machines can reason about it. Yet no essay captured some of the mood at the time better than Hilary Putnam's "Much Ado About Not Very Much." Putnam questioned the Dædalus issue itself: "Why a whole issue of Dædalus? Why don't we wait until AI achieves something and then have an issue?" He concluded:This volume of Dædalus is indeed the first since 1988 to be devoted to artificial intelligence. This volume does not rehash the same debates; much else has happened since, mostly as a result of the success of the machine learning approach that was being rediscovered and reimagined, as discussed in the 1988 volume. This issue aims to capture where we are in AI's development and how its growing uses impact society. The themes and concerns herein are colored by my own involvement with AI. Besides the television, films, and books that I grew up with, my interest in AI began in earnest in 1989 when, as an undergraduate at the University of Zimbabwe, I undertook a research project to model and train a neural network.9 I went on to do research on AI and robotics at Oxford. Over the years, I have been involved with researchers in academia and labs developing AI systems, studying AI's impact on the economy, tracking AI's progress, and working with others in business, policy, and labor grappling with its opportunities and challenges for society.10The authors of the twenty-five essays in this volume range from AI scientists and technologists at the frontier of many of AI's developments to social scientists at the forefront of analyzing AI's impacts on society. The volume is organized into ten sections. Half of the sections are focused on AI's development, the other half on its intersections with various aspects of society. In addition to the diversity in their topics, expertise, and vantage points, the authors bring a range of views on the possibilities, benefits, and concerns for society. I am grateful to the authors for accepting my invitation to write these essays.Before proceeding further, it may be useful to say what we mean by artificial intelligence. The headlines and increasing pervasiveness of AI and its associated technologies have led to some conflation and confusion about what exactly counts as AI. This has not been helped by the current trend-among researchers in science and the humanities, startups, established companies, and even governments-to associate anything involving not only machine learning, but data science, algorithms, robots, and automation of all sorts with AI. This could simply reflect the hype now associated with AI, but it could also be an acknowledgment of the success of the current wave of AI and its related techniques and their wide-ranging use and usefulness. I think both are true; but it has not always been like this. In the period now referred to as the AI winter, during which progress in AI did not live up to expectations, there was a reticence to associate most of what we now call AI with AI.Two types of definitions are typically given for AI. The first are those that suggest that it is the ability to artificially do what intelligent beings, usually human, can do. For example, artificial intelligence is:The human abilities invoked in such definitions include visual perception, speech recognition, the capacity to reason, solve problems, discover meaning, generalize, and learn from experience. Definitions of this type are considered by some to be limiting in their human-centricity as to what counts as intelligence and in the benchmarks for success they set for the development of AI (more on this later). The second type of definitions try to be free of human-centricity and define an intelligent agent or system, whatever its origin, makeup, or method, as:This type of definition also suggests the pursuit of goals, which could be given to the system, self-generated, or learned.13 That both types of definitions are employed throughout this volume yields insights of its own.These definitional distinctions notwithstanding, the term AI, much to the chagrin of some in the field, has come to be what cognitive and computer scientist Marvin Minsky called a "suitcase word."14 It is packed variously, depending on who you ask, with approaches for achieving intelligence, including those based on logic, probability, information and control theory, neural networks, and various other learning, inference, and planning methods, as well as their instantiations in software, hardware, and, in the case of embodied intelligence, systems that can perceive, move, and manipulate objects.Three questions cut through the discussions in this volume: 1) Where are we in AI's development? 2) What opportunities and challenges does AI pose for society? 3) How much about AI is really about us?Notions of intelligent machines date all the way back to antiquity.15 Philosophers, too, among them Hobbes, Leibnitz, and Descartes, have been dreaming about AI for a long time; Daniel Dennett suggests that Descartes may have even anticipated the Turing Test.16 The idea of computation-based machine intelligence traces to Alan Turing's invention of the universal Turing machine in the 1930s, and to the ideas of several of his contemporaries in the mid-twentieth century. But the birth of artificial intelligence as we know it and the use of the term is generally attributed to the now famed Dartmouth summer workshop of 1956. The workshop was the result of a proposal for a two-month summer project by John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon whereby "An attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves."17In their respective contributions to this volume, "From So Simple a Beginning: Species of Artificial Intelligence" and "If We Succeed," and in different but complementary ways, Nigel Shadbolt and Stuart Russell chart the key ideas and developments in AI, its periods of excitement as well as the aforementioned AI winters. The current AI spring has been underway since the 1990s, with headline-grabbing breakthroughs appearing in rapid succession over the last ten years or so: a period that Jeffrey Dean describes in the title of his essay as a "golden decade," not only for the pace of AI development but also its use in a wide range of sectors of society, as well as areas of scientific research.18 This period is best characterized by the approach to achieve artificial intelligence through learning from experience, and by the success of neural networks, deep learning, and reinforcement learning, together with methods from probability theory, as ways for machines to learn.19A brief history may be useful here: In the 1950s, there were two dominant visions of how to achieve machine intelligence. One vision was to use computers to create a logic and symbolic representation of the world and our knowledge of it and, from there, create systems that could reason about the world, thus exhibiting intelligence akin to the mind. This vision was most espoused by Allen Newell and Hebert Simon, along with Marvin Minsky and others. Closely associated with it was the "heuristic search" approach that supposed intelligence was essentially a problem of exploring a space of possibilities for answers. The second vision was inspired by the brain, rather than the mind, and sought to achieve intelligence by learning. In what became known as the connectionist approach, units called perceptrons were connected in ways inspired by the connection of neurons in the brain. At the time, this approach was most associated with Frank Rosenblatt. While there was initial excitement about both visions, the first came to dominate, and did so for decades, with some successes, including so-called expert systems.Not only did this approach benefit from championing by its advocates and plentiful funding, it came with the suggested weight of a long intellectual tradition-exemplified by Descartes, Boole, Frege, Russell, and Church, among others-that sought to manipulate symbols and to formalize and axiomatize knowledge and reasoning. It was only in the late 1980s that interest began to grow again in the second vision, largely through the work of David Rumelhart, Geoffrey Hinton, James McClelland, and others. The history of these two visions and the associated philosophical ideas are discussed in Hubert Dreyfus and Stuart Dreyfus's 1988 Dædalus essay "Making a Mind Versus Modeling the Brain: Artificial Intelligence Back at a Branchpoint."20 Since then, the approach to intelligence based on learning, the use of statistical methods, back-propagation, and training (supervised and unsupervised) has come to characterize the current dominant approach.Kevin Scott, in his essay "I Do Not Think It Means What You Think It Means: Artificial Intelligence, Cognitive Work & Scale," reminds us of the work of Ray Solomonoff and others linking information and probability theory with the idea of machines that can not only learn, but compress and potentially generalize what they learn, and the emerging realization of this in the systems now being built and those to come. The success of the machine learning approach has benefited from the boon in the availability of data to train the algorithms thanks to the growth in the use of the Internet and other applications and services. In research, the data explosion has been the result of new scientific instruments and observation platforms and data-generating breakthroughs, for example, in astronomy and in genomics. Equally important has been the co-evolution of the software and hardware used, especially chip architectures better suited to the parallel computations involved in data- and compute-intensive neural networks and other machine learning approaches, as Dean discusses.Several authors delve into progress in key subfields of AI.21 In their essay, "Searching for Computer Vision North Stars," Fei-Fei Li and Ranjay Krishna chart developments in machine vision and the creation of standard data sets such as ImageNet that could be used for benchmarking performance. In their respective essays "Human Language Understanding & Reasoning" and "The Curious Case of Commonsense Intelligence," Chris Manning and Yejin Choi discuss different eras and ideas in natural language processing, including the recent emergence of large language models comprising hundreds of billions of parameters and that use transformer architectures and self-supervised learning on vast amounts of data.22 The resulting pretrained models are impressive in their capacity to take natural language prompts for which they have not been trained specifically and generate human-like outputs, not only in natural language, but also images, software code, and more, as Mira Murati discusses and illustrates in "Language & Coding Creativity." Some have started to refer to these large language models as foundational models in that once they are trained, they are adaptable to a wide range of tasks and outputs.23 But despite their unexpected performance, these large language models are still early in their development and have many shortcomings and limitations that are highlighted in this volume and elsewhere, including by some of their developers.24In "The Machines from Our Future," Daniela Rus discusses the progress in robotic systems, including advances in the underlying technologies, as well as in their integrated design that enables them to operate in the physical world. She highlights the limitations in the "industrial" approaches used thus far and suggests new ways of conceptualizing robots that draw on insights from biological systems. In robotics, as in AI more generally, there has always been a tension as to whether to copy or simply draw inspiration from how humans and other biological organisms achieve intelligent behavior. Elsewhere, AI researcher Demis Hassabis and colleagues have explored how neuroscience and AI learn from and inspire each other, although so far more in one than the other, as and have the success of the current approaches to AI, there are still many shortcomings and as well as problems in It is useful to on one such as when AI does not as or or or that can to or when it on or information about the world, or when it has such as of all of which can to a of public shortcomings have captured the attention of the wider public and as well as among there is an on AI and In recent years, there has been a of to principles and approaches to AI, as well as involving and such as the on AI, that to best important has been the of with to and - in the and developing AI in both and as has been well in recent This is an important in its own but also with to the of the resulting AI and, in its intersections with more the other there are limitations and problems associated with the that AI is not capable of if could to more more or more general AI. In their Turing deep learning and Geoffrey took of where deep learning and highlighted its current such as the with In the case of natural language processing, Manning and Choi the challenges in and despite the of large language Elsewhere, and have the notion that large language models do anything learning, or In & of in a and discuss the problems in systems, the as how to reason about other their systems, and well as challenges in both and especially when the include both humans and Elsewhere, and others a useful of the problems in there is a growing among many that we do not have for the of AI systems, especially as they become more capable and the of use although AI and its related techniques are to be powerful tools for research in science, as examples in this volume and recent examples in which AI not only help results but also by design and become what some have AI to science and and to and challenges for the possibility that more powerful AI could to new in science, as well as progress in some of challenges and has long been a key for many at the frontier of AI research to more capable the of each of AI, the of more general problems that to the possibility of more capable AI learning, reasoning, of and and of these and other problems that could to more capable systems the of whether current characterized by deep learning, the of and and more foundational and and reinforcement or whether different approaches are in such as cognitive agent approaches or or based on logic and probability theory, to name a few. whether and what of approaches be the AI is but many the current along with of and learning architectures have to their about the of the current approaches is associated with the of whether artificial general intelligence can be and if how and Artificial general intelligence is in to what is called that AI and for tasks and goals, such as The development of on the other aims for more powerful AI - at as powerful as is generally to problem or and, in some the capacity to and improve as well as set and its own and the of and when will be is a for most that its achievement have and as is often in and such as A through and The to Ex and it is or there is growing among many at the frontier of AI research that we for the possibility of powerful with to and and with humans, its and use, and the possibility that of could and that we these into how we approach the development of of the research and development, and in AI is of the AI and in its what Nigel Shadbolt the of AI. This is given the for useful and applications and the for in sectors of the However, a few have made the development of their the most of these are and each of which has demonstrated results of increasing still a long way from the most discussed impact of AI and automation is on and the future of This is not In in the of the excitement about AI and and concerns about their impact on a on and the was that such technologies were important for growth and and "the that but not Most recent of this including those I have been involved have and that over time, more are than are that it is the and the and the of will the In their essay AI & and John discuss these for work and further, in & the of & to discuss the with to and and as well as the opportunities that are especially in developing In "The Turing The & of Artificial Intelligence," discusses how the use of human benchmarks in the development of AI the of AI that rather than human He that the AI's development will take in this and resulting for will on the for companies, and a that the that more will be than too much from of the and does not far enough into the future and at what AI will be capable The for AI could from of that in the is and labor and ability to are and and until automation has mostly physical and but that AI will be on more cognitive and tasks based on and, if early examples are even tasks are not of the In other are now in the world machines that that learn and that their ability to do these is to a range of problems they can will be with the range to which the human has been This was and Allen Newell in that this time could be different usually two that new labor will in which will by other humans for their own even when machines may be capable of these as well as or even better than The other is that AI will create so much and all without the for human and the of will be to for when that will the that once the first time since his creation will be with his his to use his from how to the which science and interest will have for to live and and However, most researchers that we are not to a future in which the of will and that until then, there are other and that be in the labor now and in the such as and other and how humans work increasingly capable that and John and discuss in this are not the only of the by AI. Russell a of the potentially from artificial general intelligence, once a of or ten But even we to general-purpose AI, the opportunities for companies and, for the and growth as well as from AI and its related technologies are more than to pursuit and by companies and in the development, and use of AI. At the many the is it is generally that is a in AI, as by its growth in AI research, and as highlighted in several will have for companies and given the of such technologies as discussed by and others the may in the way of approaches to AI and (such as whether they are companies or as and have have the to to in AI. The role of AI in intelligence, systems, autonomous even and other of increasingly In &

In the United States, colorectal cancer is the second leading cause of cancer-related death, with over 52,900 deaths in 2025 alone, according to the American Cancer Society. Early detection of malignant Kudo pit patterns through colonoscopic imaging is crucial for reducing mortality, as identifying these polyps and pits enables timely cancer diagnosis using direct visualization of the gastrointestinal system. Conventional diagnostic techniques for processing colonoscopies are explainable but inefficient, resulting in delays in timely treatment. State-of-the-art AI approaches that rely on image classification are hard to explain and mainly focus on a specific modality, leading to limitations in their diagnostic capabilities. For all these reasons, there exists a significant gap between AI advances and usage in a clinical setting. We bridge this significant gap between AI development and real-life diagnostic decision-making with two main contributions: (1) developing Endo-Insight Gen, a multimodal AI model that generates real-time, textual descriptions of endoscopic images, and (2) creating MED-X, a multi-agent AI system which integrates models such as Endo-Insight Gen to analyze an endoscopic image by collaborating with other models to provide a reason-based decision for Kudo pattern classification. Trained on a subset of the HyperKvasir dataset with approximately 10,000 labeled images, Endo Insight-Gen transforms visual features into clinically relevant textual descriptions. It simultaneously processes text and image inputs, integrating natural language processing and visual recognition to offer robust support for clinical and research applications. We benchmarked Endo-InsightGen against expert annotations and outputs from ChatGPT-4, LLaVA-1.5, and LLaVA-Med, with endoscopy specialists providing additional evaluation. The model was incorporated into MED-X alongside base models such as LLaVA-Med and fine-tuned models like LLaVA-Endo trained on ∼5,000 Kudo images. To overcome limited labeled data, we developed a web-based, human-in-the-loop annotation platform. A few-shot vision model generates preliminary labels, which are refined and validated by multiple experts, producing a consensus-driven dataset efficiently compared to manual labeling. MED-X uses multiple AI agents to collaboratively analyze each colonoscopic image, producing diagnostic conclusions on kudo pit pattern. Its framework is both interpretable and efficient: reasoning models and multi-agent summarization allow the system to mimic human expert panels. Endo-InsightGen shows strong concordance with expert annotations, demonstrating its potential as an interpretable clinical tool. MED-X further exhibits advanced reasoning for Kudo classification, outperforming current AI systems. Collectively, these contributions establish MED-X as an explainable diagnostic assistant, capable of accurate, efficient, and clinically aligned detection of colorectal polyps through colonoscopic imaging. Citation Format: Kushal Virupakshappa, Sowmya Sankaran, Yue Hu, Oladimeji Macaulay, Ala Jararaweh, David Arredondo, Gulshan Parasher, Avinash D. Sahu. Explainable AI with multi-agent collaborative system for colonoscopic polyp detection and Kudo pit classification [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Cancer Evolution: The Dynamics of Progression and Persistence; 2025 Dec 4-6; Albuquerque, NM. Philadelphia (PA): AACR; Cancer Res 2025;85(23_Suppl):Abstract nr A032.

Background: Large language model (LLM)-based agentic systems are evolving beyond single-turn generators into autonomous, toolusing, multi-agent workflows with persistent memory and self-directed planning. When these agents collaborate, hallucinations no longer remain local; they can propagate across agent boundaries and trigger real-world operational failures.  Objective: This paper provides a structured foundation for building secure, reliable, and hallucination-resilient agentic AI by surveying failure modes and proposing a layered trust taxonomy with enforceable controls.  Methods: We reviewed literature on LLM reasoning, retrieval-augmented generation (RAG), hallucination detection, multi-agent frameworks, and AI governance (including zero-trust security principles) and catalogued failure modes across reliability and security dimensions.  Results: We present a seven-layer trust taxonomy spanning identity, planning, communication, memory, retrieval, execution, and oversight. From this taxonomy, we derive six reusable secure-coordination design patterns and propose a model-agnostic reference architecture for auditable, policy-enforced agentic workflows.  Conclusion: Trustworthiness in agentic AI is fundamentally a system property, not merely a model property. The proposed taxonomy and design patterns provide practical, implementation-independent guidance for securing multi-agent LLM deployments in research and high-assurance enterprise contexts.  Plain Language Summary: AI systems built from large language models can now plan tasks, use tools, and work together as teams of specialized agents. This collaboration creates new dangers: when one agent fabricates information, the other agents may act on it as though it were true, spreading errors throughout the system. This paper maps where trust breaks down, from agent identity and message passing to memory and tool use, and proposes design rules and a system blueprint for more reliable and secure AI teams. agentic AI, multi-agent systems, trustworthy AI, hallucination mitigation, retrieval-augmented generation, tool use, zero trust, AI governance.

There have been a number of studies on efficient planning in the MAS context. For example, GPGP (Decker &amp; Lesser, 1992) is a general framework for generating effective plans using task and resource relationships among agents. Our method can be used in this framework to identify which abstract plan (task) should be refined first so that the map of the task relationships related to the plan can be created. Hierarchical planning and coordination issues for improving MAS planning have also been discussed. For example, Ref. (Clement et al., 2001) proposed choosing the most appropriate abstract task/plan on the basis of summary information derived from the primitive tasks and plans in a bottom-up fashion. This method can avoid hopeless planning if some resources are recognized to be insufficient at an abstract level. It also introduced fewestthreats-first (FTF) heuristics to choose a lower (deeper) plan. Our approach focuses on the cases where conflicts can be accurately identified at only deeper levels, because the tasks, resources, and their environment in an abstract model are described in an abstract way. Furthermore, a plan with fewer conflicts does not always lead to a better plan; it is possible that only one conflict fails to be resolved but that conflict is nonetheless a critical one. The idea behind our research is that, although conflicts may be invisible at abstract levels (including the SL), there is a tendency that conflicts often occur depending on the environmental factors related to the availability and use of resources, such as the location of agents, the kind of resources, and type of agents, as well as on the kind of task. Hence, we aim at expressing and distinguishing these situations by using CPs in order to enable agents to statistically learn the difficulty of conflict resolution and the quality of a resulting plan. A number of issues related to MAS planning have been investigated in case-based reasoning (CBR) or its related domains. For example, (Giampapa &amp; Sycara, 2001) proposed a conversational case-based reasoner, called NaCoDAE, which is a type of agent in their MAS applications and helps users decide a course of action by engaging them in a dialogue in which they must describe the problem or situation of assigning missions to platoons. Plan reuse for the same/similar situations in a MAS context has also been proposed for MAS coordination (Sugawara, 1995) and collaboration (Plaza, 2005). A remarkable work similar to our approach is (Macedo &amp; Cardoso, 2004), where a case is used to expand an abstract plan to a less abstract one in HTN, although we focus on avoiding conflicts and/or selecting costless conflicts. In this sense, our motivation is more similar to that in (Aha et al., 2005) which applied CBR to a real-time strategy game. Our work is also related to hierarchical reinforcement learning, such as (Dietterich, 1998; Kaelbling, 1993; Sutton et al., 1998), because an abstract task is considered to be a subroutine or a subfunction to be learned. For example, in the MAXQ approach (Dietterich, 1998), a task is divided into subroutines that are individually learned by RL methods. Our approach is to select an appropriate subroutine for each situation. In MAXQ, the conflict discount is assumed to have been learned at lower levels. However, in a multi-agent setting, it is naturally difficult to define the task hierarchy for all agents simultaneously. One clear limitation of our method is that the reliability of cd values heavily depends on the accuracy of the SL conflict detection and time-estimation processes. Thus, it is very important to select the appropriate SL and carefully describe the SL model. For example, if level 1 in Figure 1 is the SL, our method does not work well since that level is too abstract. As mentioned above, another issue is that the use of optional data in CPs is important for distinguishing one situation from another. To distinguish situations, our method needs the

This chapter addresses whether AI can understand me. A framework for regulating AI systems that draws on Strawson's moral philosophy and concepts drawn from jurisprudence and theories on regulation is used. This chapter proposes that, as AI algorithms increasingly draw inferences following repeated exposure to big datasets, they have become more sophisticated and rival human reasoning. Their regulation requires that AI systems have agency and are subject to the rulings of courts. Humans sponsor the AI systems for registration with regulatory agencies. This enables judges to make moral culpability decisions by taking the AI system's explanation into account along with the full social context of the misdemeanor. The proposed approach might facilitate the research and development of intelligent analytics, intelligent big data analytics, multiagent systems, artificial intelligence, and data science.