In adversarial video games, human players and non-player characters (NPCs) compete to win certain goods or targets. To gain an advantage and create realistic behaviors, NPCs could cooperate by forming groups or coalitions. Surprisingly enough, they typically do not. Consequently, we have concentrated our efforts on providing game developers with the necessary tools to enhance virtual characters with teamwork capabilities. To start with, we address this challenge by observing that a classical Reconfiguration Algorithm, known in MAS, can serve as an umbrella for teamwork-related activities of NPCs. Reconfiguration amounts to intelligent replanning in response to changing circumstances. By analogy, our solution, inspired by the reconfiguration process, takes the control over the game and keeps it going by utilizing mechanisms that facilitate cooperation under adversarial conditions. As a proof of concept, we have developed, implemented, and tested Ara, the Adversarial Reconfiguration Algorithm, alongside an encompassing agent-oriented framework called Arag and a simple stealth video game called Treasure Hunt. Arag facilitates a flexible approach to aspects of teamwork such as belief representation, communication, reconfiguration, and task allocation. As demonstrated by Treasure Hunt, NPCs coordinated by Arag outperform those lacking cooperative capabilities, even with the adopted simplifying design choices.
Achieving expert-level performance through simulation-based training relies heavily on complex and adaptable scenarios, but creating these scenarios manually is often laborious and resource-intensive. Large Language Models (LLMs) offer a promising avenue to automate and enhance scenario generation. However, their reliance on purely sequential text generation in standard prompting settings can hinder consistent understanding in complex systems. We present a multi-agent reasoning framework to leverage LLMs for puzzle generation within the 2D Physics Puzzle Environment CREATE (Chain REAction Tool Environment) to overcome these limitations. This testbed is used as a simplified analogy for scenario generation to allow the development of fundamental LLM capabilities needed to collaboratively design and solve intricate challenges, with a long-term goal of application in domains such as military training. Our framework employs a multi-agent ReAct architecture, integrating reasoning and action feedback loops to dynamically interact with CREATE. By assigning distinct roles, such as solver and designer, to individual agents, our framework preserves the complex reasoning pathways required for solving and generating puzzles—enabling complex reasoning that was too difficult to achieve with basic prompting or single-agent approaches. This work represents a step towards more robust LLM-driven scenario generation by demonstrating the ability of a multi-agent system built on our framework, while interacting with CREATE simulations, to collaboratively perform multi-step reasoning and adapt to environmental constraints. While not yet achieving real-world scenario generation, our findings demonstrate the potential of LLMs to generate solvable puzzles aligned with user prompts. However, we also highlight and address persistent challenges with their reasoning about precise spatial relationships and understanding complex, multi-step chain reactions, which are crucial for generating more advanced scenarios. We conclude by discussing the future role of multi-agent LLM frameworks in creating realistic and adaptable training scenarios for various applications, building upon the foundational capabilities developed in this work. Examples and our full code are available at: https://github.com/binzeli/Puzzle_Generation .
The connections between neuroscience findings and Artificial Intelligence (AI) are very strong, but literature that analyzes how neuroscience inspired AI, and viceversa, mainly takes a machine learning point of view. However, intelligent software agents modeled after the Belief-Desire-Intention (BDI) architecture have many ties with neuroscience. Some are explicitly expressed, others are less evident and deserve to be better addressed and understood. In order to explore such ties and make their hidden potential exploitable, we introduce the BDI-inspired MEDiTATe conceptual framework encompassing theory of Mind, Emotions, Deep TAlk, and small Talk. MEDiTATe is intended as a principled means to analyze the connections between neuroscience and BDI approaches in a systematic way, and to interact with neuro-scientists by sharing a common terminological ground. The main contribution of this paper is indeed to survey the relevant scientific literature and organize the findings of this review coherently with the MEDiTATe vision. Nonetheless, most modules of MEDiTATe have been, or may be, implemented using a well known toolkit for BDI agents, Jason. In this sense, the possibility to move MEDiTATe from the conceptual level to the practical one is backed up by existing software tools. Targeting Jason and its JaCaMo extension only is a limitation of the current MEDiTATe approach, but it is the limitation that makes a shift from theory to practice feasible. MEDiTATe features small talk and deep talk that we conjecture to be related but distinct cognitive functions, each with its own purpose and possibly dedicated different brain areas. We expect that MEDiTATe – once fully developed in a practical Jason-based toolkit – may support the study of these functions and of their relations with other, better understood, cognitive processes, possibly inspiring experiments by neuro-scientists to validate the hypothesis. In fact, in our long-term vision, MEDiTATe should offer to computer scientists and neuro-scientists a shared gym for experimenting models and theories of brain functioning.
The rapid development of LLM-based language agents has led to a proliferation of architectures described using ad hoc and inconsistent methods, making it difficult to compare, reproduce, extend or even understand them. To address this, we introduce FALAA: a framework that standardizes the description of language agent architectures through a structured set of components—Planner, Executor, Evaluator, Reflector, Memory, and Environment—and a dual-level methodology combining UML diagrams and OCL specifications. FALAA offers both conceptual clarity and formal precision, enabling unambiguous definitions of agent behaviors and responsibilities. We illustrate its effectiveness through case studies of two representative agents: Reflexion and Retroformer. The formalization reveals critical ambiguities in both architectures, such as vague component definitions and underspecified memory handling. These results show how FALAA enhances clarity, supports architecture comparison, and uncovers design limitations, making it a promising tool for the development and analysis of future language agents.
In traditional BDI (Belief-Desire-Intention) Multiagent Systems (MAS), agent communication languages such as KQML (Knowledge Query and Manipulation Language) facilitate structured message exchange and are supported by widespread BDI implementations like Jason. However, KQML lacks mechanisms to account for the situatedness of agents within dynamic environments. This paper proposes an extension to KQML, KQML-S, that incorporates the notion of locality, enabling message semantics to adapt based on the agents’ shared context. Specifically, we introduce a framework where agents within the same logical communication space, or “logical room”, perceive indirect updates from interactions occurring within their locality. We present the theoretical foundations of KQML-S and its implementation within VEsNA (Virtual Environments via Natural language Agents), a framework extending Jason with chatbots for natural language interaction, and a Virtual Reality environment implemented in Godot. KQML-S bridges two foundational features of agents: social ability and situatedness. VEsNA agents, being natively situated in Virtual Reality and inheriting KQML communication from Jason, fully exploit the potential of KQML-S and showcase its usefulness in those scenarios where agents are logically or physically embodied in a discrete, “room-based” environment.
A social machine is a Web application that enables users to interact flexibly and creatively to carry out social processes. Currently, social machines are realized via procedural technologies such as Web services. These approaches do not capture the social semantics at the heart of a social machine. Capturing the semantics of social processes would be crucial to enhancing user autonomy, accountability, interoperability, and decentralization. We present Fluid, a decentralized multiagent architecture in which the semantics of a Web application is represented foremost as a social protocol that captures the applicable norms. Unlike data decentralization architectures such as Solid, Fluid decentralizes not just the data, but also the application logic. Our contributions are the following. One, we demonstrate how Fluid promotes user autonomy and introduces accountability as a counterbalance to autonomy. Two, we demonstrate how interesting sociotechnical patterns, e.g., relating to information governance may be captured in Fluid. Three, we demonstrate how Fluid applications may be realized using data decentralization technologies such as Solid.
This paper proposes a structured approach to integrating Large Language Models (LLMs) into Multi-Agent Systems (MAS) by revisiting and extending the fundamental Agent-Oriented Software Engineering (AOSE) concept of “roles.” Traditional AOSE methodologies provide well-defined processes for modeling agents, roles, goals, and interactions, yet contemporary LLM-based MAS frameworks typically lack such systematic engineering foundations. We highlight how ad hoc development practices in LLM-enhanced MAS—often driven by prompt engineering or role-playing strategies—can lead to inconsistencies and reduced maintainability. Through a critical examination of role definition, specification, and implementation, we identify several gaps in terms of software engineering. To bridge these gaps, we propose a hybrid role-based architecture where we treat roles as first-class entities at run-time encapsulating both traditional AOSE design principles and LLM-driven functionalities. By laying this groundwork, we aim to foster more robust, scalable, and transparent engineering of LLM-enhanced MAS.
This paper explores the semantics of Tumato 2.0, a constraint-based planning framework, through the lens of Linear Temporal Logic (LTL). Tumato enables the generation of policies for autonomous agents, ensuring safe and robust goal-oriented behavior. The framework guarantees that critical safety constraints hold across all potential outcomes of non-deterministic actions, while pre-computed policies eliminate the need for runtime decision-making. By translating Tumato’s language constructs into LTL, we formalize its approach to handling safety, liveness, and robustness properties. This contribution offers a foundation for reliable agent behavior under real-world uncertainties, as well as improved interpretability. We further demonstrate the semantics of Tumato’s specification language through a case study, demonstrating how LTL guides system specification and supports potential formal verification efforts. These contributions align with key challenges in engineering intelligent and multi-agent systems, focusing on safety, correctness, and robust operation within complex environments. Overall, this work emphasizes the importance of declarative approaches in delivering reliable solutions for real-world applications.
The rapid expansion of sensor-based networks introduces major challenges in scalability, adaptability, and knowledge transfer, especially in open environments where new subsystems can dynamically join or leave. In this work, we propose a Holonic Active Distillation architecture within a Holonic Multi-Agent System (HMAS) to address these issues. Our approach integrates Clustered Stream-Based Active Distillation (CSBAD), a framework in which specialized student models collect local data, query pseudo-labels from teacher models, and cluster into groups of similar sensors. Results show that the holonic organization balances local specialization with global generalization, while efficiently adapting to sensor departures and re-integrations. We also analyzed trade-offs among incremental model updates, system reorganization, and scalability limits. Our findings highlight the advantages of holonic learning for multi-sensor systems while identifying key challenges related to model drift and long-term adaptation.
This paper presents the outcomes of an open-floor session held at the 13th International Workshop on Engineering Multi-Agent Systems (EMAS 2025), aimed at co-developing a research roadmap for the EMAS community. Participants collaboratively identified and prioritised challenges in engineering large-scale, adaptive multiagent systems, particularly considering the need to engineer systems that can seamlessly integrate learning and reasoning. Through structured group discussions, four key challenges emerged: explainability in heterogeneous environments, environment modeling, handling dynamic contexts, and communication standardisation. For each of the challenges, participants proposed and ranked potential solutions based on impact and effort. The resulting roadmap highlights concrete research directions toward engineering intelligent, explainable, and interoperable multiagent systems that effectively integrate reasoning and learning in dynamic environments.
Despite the significant advances in subsymbolic artificial intelligence over the last decade, including novel Large Language Model (LLM) based methods, this technology comes with high, even prohibitive, development and application costs. Recent research suggests that instead of blindly increasing the model sizes and deploying larger numbers of better accelerators, there are benefits from a focus on the model structures and methods employed instead. An approach that has gained in popularity lately that addresses this is neurosymbolic reasoning. By focusing on various ways of combining subsymbolic, mostly neural network-based, computations with classical symbolic reasoning, it promises to alleviate the computation demands of pure deep learning approaches by guiding the learning process with symbolic knowledge. In this paper, we introduce the Modular Hybrid Agent Architecture (MHAgentA), a cognitive agent architecture, along with a Python framework that implements it. This architecture is designed to facilitate the prototyping and deployment of neurosymbolic agents that follow Kahneman’s System 1, System 2 model. We provide a breakdown of a high-level view of the architecture and outline the technical details of its implementation.
Explainable agency (XAg) aims at providing users with insights about the reasoning and decisions taken by an agent. Most of the newer XAg approaches are particularly useful as explanations for developers and researchers. In contrast, the novel XAg framework presented in this paper aims to address the explanation needs of end users, including both domain experts and lay users. It is a challenging task since this kind of users is not familiar with the methodological and technical aspects of agency. We propose a representation for end user questions and potential explanatory answers in both a verbal and a formal description as well as a mapping structure of questions to multiple possible explanations. We develop a pattern-based approach to extract explanatory content from an execution log and to validate potential answers to a user question which is based on the TriQPAN decision patterns from the literature [14]. We organize the novel concepts in a four-layered architecture with layers for end user questions, validation logic, TriQPAN patterns, and answer text generation. A running sample from a Jadex-BDI project on autonomous mobility on demand provides a demonstration scenario to illustrate some data structures and pseudocode. Further, it highlights the plausibility of our novel XAg framework.
Extracting alphanumeric data from form-like documents such as invoices, purchase orders, bills, and financial documents is often performed via vision (OCR) and learning algorithms or monolithic pipelines with limited potential for systemic improvements. We propose an agentic AI system that leverages Large Language Model (LLM) agents and a reinforcement learning (RL) driver agent to automate consistent, self-improving extraction under LLM inference uncertainty. Our work highlights the limitations of monolithic LLM-based extraction and introduces a modular, multi-agent framework with task-specific prompts and an RL policy of rewards and penalties to guide a meta-prompting agent to learn from past errors and improve prompt-based actor agents. This self-corrective adaptive system handles diverse documents, file formats, layouts, and LLMs, aiming to automate accurate information extraction without the need for human intervention. Results as reported on two benchmark datasets of SOIRE, and CORD, are promising for the agentic AI framework.
AI agents have become increasingly significant in various domains, enabling autonomous decision-making and problem-solving. To function effectively, these agents require a planning process that determines the best course of action and then executes the planned actions. In this paper, we present an efficient on-device Planner-Action framework that separates planning and action execution into two components: a planner agent, or Octo-planner, optimized for edge devices, and an action agent using the Octopus model for function execution. Octo-planner first responds to user queries by decomposing tasks into a sequence of sub-steps, which are then executed by the Octopus action agent. To optimize performance on resource-constrained devices, we employ model fine-tuning instead of in-context learning, reducing computational costs and energy consumption while improving response times. Our approach involves using GPT-4 to generate diverse planning queries and responses based on available functions, with subsequent validations to ensure data quality. We fine-tune the Phi-3 Mini model on this curated dataset, achieving a 97 NexaAI Hugging Face repo . For its performance on mobile devices, please refer to our YouTube video demo .
Explanation is important to supporting appropriate levels of trust in autonomous systems. However, work in XAI (eXplainable AI) is focused on explanation of single system components, such as a machine learning algorithm or decision-making module. This paper: (1) argues that we need to develop ways to engineer explainable systems consisting of multiple components, and identifies this as a challenge for the community; (2) proposes an approach for explaining multi-component autonomous systems; (3) identifies integration issues that need to be addressed to make this vision a reality; and (4) poses a number of research challenges and questions that need to be addressed.
Domain-Driven Design (DDD) emerged in the last two decades as an effective approach adopted especially in agile software development to tackle "the complexity at the heart of software", to quote one of its main mottos. In this paper, we are interested in exploring the bi-directional conceptual interaction between DDD and Agent-Oriented Software Engineering (AOSE)-and the synergies stemming from their fruitful integration.
Agent programming languages and their interpreters are crucial in autonomous decision-making. While formal methods are extensively utilized to ensure the correctness of agent programs, their application for verifying the implementation correctness of interpreters remains infrequent. To formally specify and verify autonomous decision-making, we proposed vGOAL and implemented its interpreter. The implementation correctness of the vGOAL interpreter is crucial for users to gain trust in the vGOAL approach. Using program verification is one option, yet this would require a huge effort to verify the correctness of the vGOAL interpreter. In this paper, we propose integrating an SAT-solving component into the vGOAL interpreter to enhance confidence in its core component: minimal model generation. The SAT-solving component consists of two subcomponents: an SAT-encoding component and an SAT solver. Leveraging PySAT for its interface to advanced solvers, our main contribution lies in the SAT encoding. We devise an algorithm to encode the inputs and outputs of the core component into a satisfiable CNF formula. Importantly, we justify that this algorithm generates a satisfiable CNF formula only if the result is correct. We demonstrate the practicality and efficiency of this SAT-solving approach using a case study involving an autonomous transportation system with three mobile robots.
Innovative solutions to societal challenges require the generation of creative ideas. Collaboration between humans and multi-agent systems (MAS) is a promising approach for idea generation, yet fostering creative discussions remains a challenge. This paper proposes Synergistic Platform for Advancing and Reinforcing Knowledge through Interactive Tools (SPARKIT) and SPARK-flow leveraging mind maps to facilitate idea generation within a MAS framework. SPARKIT supports idea generation with two types of large language model (LLM)-based agents: Debater Agents and the Moderator Agent. Debater Agents offer varied perspectives from their expertise, while the Moderator Agent structures discussions into a mind map to enhance user-agent collaboration. SPARK-flow is designed to stimulate creative idea generation by orchestrating discussions among these agents. A distinctive aspect of SPARKIT and SPARK-flow is that the agents facilitate the discussion and its structuring into a mind map, reducing the user's burden compared to existing methods. This paper compares different methods of discussion using language models and reveals that the impact of the discussion method on the creativity of ideas is consistent between humans and language models, with mind maps significantly enhancing creativity.
Multi-agent Reinforcement Learning (MARL) has shown significant success in solving large-scale complex decision-making problems in multi-agent systems (MAS) while facing the challenge of increasing computational cost and training time. MARL algorithms often require sufficient environment exploration to achieve good performance, especially for complex environments, where the interaction frequency and synchronous training scheme can severely limit the overall speed. Most existing RL training frameworks, which utilize distributed training for acceleration, focus on simple single-agent settings and are not scalable to extend to large-scale MARL scenarios. To address this problem, we introduce a Scalable Asynchronous Distributed Multi-Agent RL training framework called SADMA, which modularizes the training process and executes the modules in an asynchronous and distributed manner for efficient training. Our framework is powerfully scalable and provides an efficient solution for distributed training of multi-agent reinforcement learning in large-scale complex environments. Code is available at https://github.com/sadmaenv/sadma.
This paper explores the integration of the Gaia-X framework into a Multi-Agent System (MAS) for managing a smart freight fleet, emphasizing identity and trust management. Focusing on a subsystem of delivery agents and autonomous robots, this study exemplarily illustrates how Gaia-X can be integrated in existing ecosystems consisting of software agents and appertaining services and assets. By utilizing Organizational Credential Managers (OCMs), mediator services and wallets, the paper addresses the challenges of credential management and connectivity for mobile edge devices like delivery robots. This integration showcases the potential of Gaia-X to improve the security and interoperability of smart freight systems, contributing to the advancement of trusted digital ecosystems in the logistics sector.
