The integration of Belief-Desire-Intention (BDI) agents with robotic systems offers a promising approach to programming autonomy and decision-making in complex environments. This paper extends an interface to combine the Jason agent programming language with the Robot Operating System (ROS) to control multiple robots. We evaluate this implementation through three case studies of increasing complexity and benchmark it against a functionally equivalent Python-based implementation using qualitative and quantitative metrics. The outcomes are directly compared with those of the Python baseline. Our study outlines the experimental design and evaluation methodology, aiming to clarify the trade-offs between agent-oriented and imperative approaches in multi-robot systems and to identify scenarios in which BDI-based models provide advantages in responsiveness, robustness, and scalability.
Large-language models (LLMs) incorporated into agentic systems present exciting opportunities to accelerate drug discovery. In this study, we examine the modularity of LLM-based agentic systems for drug discovery, i.e., whether parts of the system are interchangeable, a topic that has received limited attention in drug discovery. We compare the performance of different LLMs and the effectiveness of tool-calling agents versus code-generating agents. Our case study, comparing performance in orchestrating tools for chemistry and drug discovery using an LLM-as-a-judge score, shows that Claude-3.5-Sonnet, Claude-3.7-Sonnet and GPT-4o outperform alternative LLMs such as Llama-3.1-8B, Llama-3.1-70B, GPT-3.5-Turbo, and Nova-Micro. Although we confirm that code-generating agents outperform the tool-calling ones on average, we show that this is highly question- and model-dependent. Furthermore, the impact of replacing the system promptl is dependent on the question and model, underscoring that even in this particular domain one cannot replace components of the system without re-engineering. Our study highlights the necessity of further research into the modularity of LLM-based agentic systems to enable the development of reliable and modular solutions for real-world problems.
Business processes specify patterns of activities expected to occur in an organization to achieve a business goal. However, actual executions may deviate from the expected ones and not necessarily because of mistakes, but also due to adaptation to circumstances or circumvention of limitations. To cope with this issue, organizations rely on accountability frameworks to get reliable information about deviations to undertake corrective measures for their processes, goals, or organization. However, to be effectively carried out, accountability has to rely on additional structures to be put in place, such as approvals and checking activities, which come at a price. So, these additional structures should be parsimoniously introduced. The paper presents a strategy to identify the activities that are riskier than others, and for which it is reasonable to foresee accountability structures. This evaluation is based on execution logs, analyzing which activities are not executed as expected from the business process.
Achieving high realism in Virtual Reality (VR) simulations with many Virtual Humans (VH) is a complex task and involves many different and challenging aspects. This paper focuses on enhancing VH autonomy to enable intelligent behaviours and interaction with the environment. Drawing from research in autonomous intelligent agents and Multiagent Systems (MAS), we analyse which results that the MAS community already obtained can be used in VR to improve VH behaviour, and which requirements should simulation tools fulfill in order to support them. To showcase our vision, we present the JaCa-Unity library for bridging the JaCaMo platform and Unity, and we show how JaCa-Unity has been integrated in VEsNA , an ecosystem that allows Jason agents to interact with humans via a chatbot-like interface and to be immersed within a VR environment.
We propose the study of diffusion of two secrets in a heterogeneous environment, that is, there are two directed networks over the same set of agents. Our model combines two group identification processes, each with independent desiderata. Specifically, we have two types of targets: constructive, where we aim to expose a given group of agents to a secret, and destructive, where we aim to prevent a given group of agents from being exposed to a secret. To secure these targets, we can employ different control actions: either deleting an agent or introducing a previously latent agent. For all combinations of the aforementioned targets and actions, we provide a complexity classification. Since classical analysis mostly yields negative results, we adopt a parameterized perspective, considering (combinations of) natural parameters, such as the number of controlled agents or the number of rounds of the activation process. Taking a parameterized perspective, we provide several FPT and XP algorithms along with their respective lower bounds.
This paper proposes an AgentSpeak plan generation system which is capable of handling non-deterministic actions and partially observable environments. It utilizes a pipeline transforming the initial specifications of the planning problem into the inputs for state-of-the art Fully Observable Non-Deterministic (FOND) and Partially Observable Non-Deterministic (POND) planners, whose outputs are translated into AgentSpeak plans.
We address the problem of assigning houses to incoming agents, assuming that some of the houses are already inhabited. We are given a city’s topology (modelled as an undirected graph), a set of inhabitants occupying some of the vertices (or houses) together with their preferences, and a set of newcomers we need to accommodate in the unoccupied houses. Within this model, we aim to find housing where no agent can benefit from swapping a house with another agent or jumping to an unoccupied house. Our contribution is three-fold. First, we study under which conditions a stable housing is guaranteed to exist. Second, we provide a complexity analysis of the associated computational problems. Apart from showing that the problems we consider are computationally hard in general, we introduce several tractable algorithms for several special cases. Finally, we investigate the economic efficiency of stable housing through the lens of (utilitarian) social welfare and analyse the Price of Stability and the Price of Anarchy in our setting.
Agent development and deployment frameworks have seen little development recently, and by far the most popular framework is Jade, with SPADE coming strong for Python deployments. Flash-mas – the Flexible, Lightweight Agent Shell – is a framework for the deployment of multi-agent systems comprising a variety of entities besides agents, with a great level of modularity. In Flash-mas, any part of the system – agents, but also nodes, communication infrastructures, services, and agent components – can have an arbitrary implementation and is loaded dynamically when a scenario is started. Moreover, any other type of entity is automatically, dynamically integrated, and can easily interact with other entities on the same node or on other nodes. We present in this demonstration our current progress in the development of Flash-mas, how new types of entities can be seamlessly integrated in a deployment, how machine learning models written in Python can integrate with Java agents, how remote visualization is possible via a web application, and how monitoring the progress of a scenario is possible thanks to log summarizers and finite state machines.
Modeling pedestrian motion has critical applications in public safety, urban planning, and infrastructure design. Understanding how pedestrians behave in real-world environments is essential for predicting and controlling collective movements, which can become hazardous and lead to dangerous situations. When focusing on the mutual interactions among pedestrians (agents), their motion can be modeled through a system of ordinary differential equations (ODEs) that describe the agents’ dynamics at the microscopic level. In this approach, the parameters used in the ODEs system play a crucial role in accurately reproducing walking behavior under realistic conditions. Typically, these parameters are either assumed based on values from the literature or estimated from available datasets. In this work, we present a hybrid theoretical–data-driven framework for modeling pedestrian motion at the microscopic level. Specifically, we employ the social-force model and estimate its key parameters through a data-driven optimization procedure informed by experimental data. Furthermore, we introduce a stochastic differential equations (SDEs) formulation of the model, which enhances its ability to capture random effects in human movement. Finally, we analyze the influence of environmental constraints, highlighting the potential for implementing pedestrian motion control measures.
In this article, we investigate the explainability of the fair roommate matching problem, where pairs of agents must be formed based on their ordinal preferences and where the existence of a fair matching is not guaranteed. We focus on explaining instances where rank-envy-free matchings—a relaxation of envy-freeness—do not exist. Our goal is to provide short and easy-to-understand explanations for users. To this end, we propose three formulations of our roommate problem as instances of the Boolean satisfiability problem (SAT) and compare their properties. We then generate explanations derived from Minimal Unsatisfiable Subsets (MUSes) for all formulations by adapting an existing explainability procedure recently introduced in the context of a fair division problem. Our adaptation uses a sequential algorithm to avoid redundant arguments in the explanation that can happen in the initial procedure, but its efficiency depends on chosen arguments at each step. These choices are guided by proposed heuristics. We conduct experiments to empirically evaluate our heuristics, compare our SAT formulations as well as the MUSes we choose to derive explanations.
Swarms of autonomous underwater vehicles (AUVs) offer a promising solution for complex missions in dynamic environments with limited communication. Unforeseen events, such as currents, equipment failures, or energy shortages, can prevent some AUVs from completing tasks, requiring rapid redistribution among the remaining vehicles. Decentralized coordination under such conditions is challenging, especially without reliable feedback on overall mission progress. This work investigates a decentralized decision-making process for task reallocation, grounded in opinion dynamics. A key parameter is an agent’s stubbornness, capturing the tendency to persist with an opinion despite uncertainty. Stubbornness influences when and how the swarm reassigns tasks. Simulations across cooperation strategies and swarm compositions show that allowing controlled divergence in agent opinions, rather than enforcing rapid consensus, improves mission success and resilience. Diverse viewpoints reduce groupthink, encourage exploration, and support flexible adaptation to environmental changes. These results highlight the potential of simple, decentralized heuristics like adaptive stubbornness for robust coordination in uncertain underwater environments.
Adaptive learning refers to approaches that dynamically tailor content, pacing, and feedback to an individual learner’s progress, preferences, and prior knowledge. Large Language Models (LLMs) and Artificial Intelligence (AI) agents are transforming the landscape of higher education by enabling new forms of personalized, interactive, and adaptive learning. This paper argues for the intentional integration of AI agents powered by LLM into higher education to support adaptive learning grounded in the established pedagogical framework (Bloom’s Taxonomy). We propose that such an AI system can serve not merely as automated content providers but as intelligent, dialogic partners that guide learners across cognitive levels and disciplines through real-time feedback, personalized support, and collaborative knowledge construction. Drawing from educational theory and a practical use case developed by the authors, we demonstrate how knowledge-driven Human-AI interaction can be applied to diverse learning tasks to enhance learner engagement, cognitive development, and instructional scalability. This paper advocates for a future where AI agents augment rather than replace instruction, enabling meaningful, personalized, and cognitively rich learning experiences in higher education.
An agent architecture provides a structured framework for handling the complexity of agent behaviour. It plays a central role not only in agent-oriented software development but also in agent-based simulation. Classical agent architectures elaborate the sense-reason-act cycle – yet in many scenarios, this is not sufficient for formulating realistic behaviour. An example is an agent model for simulating decision-making in the well-studied problem of human navigation. Here, human decision making is organized in interwoven sub-processes such as perception, movement, planning, and spatio-temporal reasoning. These processes occur concurrently, making human navigation inherently multi-level in nature. In this paper, we demonstrate how an agent architecture that allows modelling decision-making as concurrent processes can yield realistic and more easily modelled complex agent behaviour. Bringing together research on agent architectures, cognitive architectures, human behaviour modelling, and computational psychology, we present an example that integrates key subprocesses essential for simulating agents guided by cognitively plausible models of human navigation.
Multi-agent systems for resource allocation (MRAs) provide a framework for modelling competitive resource allocation in distributed computing. In an MRA, agents interact over discrete rounds, aiming to satisfy their allocation goals. We present a technique for synthesising collectively optimal strategies for MRAs. A collectively optimal strategy not only guarantees goal satisfaction but also maximises the frequency of goal achievement throughout the system’s run. While previous work was limited to synthesising strategies for finite runs that terminate after a fixed number of rounds, our new approach enables the synthesis of optimal strategies for infinite runs, enhancing their practical applicability. Our approach encodes the synthesis problem in propositional logic and solves it using maximum satisfiability (Max-SAT). We encode loops in the state space of an MRA that may arise when agents follow a uniform strategy. The encoding incorporates weighted soft clauses, representing instances in which an agent achieves its goal at a particular state within a loop, given an assumed loop size. Solving Max-SAT yields a truth assignment that maximises the total weight of satisfied soft clauses. In our setting, the assignment characterises a strategy that leads to a loop in which the frequency of goal achievement is collectively optimal. We implemented our approach within the SATMAS tool and conducted experiments demonstrating promising synthesis results.
Autonomous traffic agents ( ATAs )–automated systems with high level of autonomy in traffic environments–must not only guarantee safety but also act in accordance with legal, social, and moral values. In this short version, we adopt the epistemic game-theoretic conflict model of [6] to characterise value conflicts—situations where competing, value-laden goals cannot all be satisfied. As a mean to align the decision making of an ATA with stakeholder preferences, we introduce Value-Aligned Operational Design Domains (VODDs). They represent autonomous decision making scopes that guide an agent’s conflict resolution and specify handover rules.
This paper presents a proof-of-concept implementation of the BOID logic in smart scheduling tasks. We leverage the potential of using embedding models as a foundation for building knowledge bases and ‘mining’ BOID rules with priorities. With this approach, we address the problems of extracting logic rules from the data on users and conference topics. Our rule mining method addresses a key pitfall of logic-based systems: their dependence on domain experts to generate logic rules. We apply the model to the domain of conference planning, where we build a personalized conference attendance schedule for users of the BOID app. In generating the schedule, we rely on the three main resources for decision making: rule-based reasoning (BOID), embedding models and human input. Each of the resources has some specific role in eliciting either of the Belief, Obligation, Intention and Desire components. This makes our model ‘hybrid’ in at least two different ways, namely, as a hybrid learning-reasoning model and as a hybrid human-AI model. One of the main advantages of our model is interpretability through tracing the BOID rules underlying each decision in generated schedules.
AgentSpeak(AI) extends the definition of an agent provided in AgentSpeak(L) to include learning capabilities which are integrated into the AgentSpeak(L) interpreter, and allow the agent to acquire new plans at runtime. The extension leverages agent beliefs, and introduces the concept of learning processes which can be linked to a triggering event at the agent level using a learning process template. Agents can control the learning process, amend the configuration of the process during training, and deliberate over the process by examining metrics, encapsulated in beliefs. With these controls, agents can form deliberate learning strategies which operate over innate learning processes. A high level overview is given of the extension and relevant changes to the AgentSpeak(L) interpreter, plus a demonstration of the extension in the AgentSpeak(L)-based programming language ASTRA.
We study the model called Agents to Facilities with Bans Problem (AFB), which belongs to the intersection of algorithmic game theory and stable matching theory: There are m facilities and n agents, each agent has a priority list of the facilities, each facility has a limited capacity and may contain a list of banned agents. We design a mechanism that finds Pareto optimal, minimal envy-2 solution to AFB. We prove that this mechanism is strategyproof and runs in time O(nm).
The ability for agents to explain the reasons behind their decisions and behaviour is fundamental for human developers or participants to build trust in a system, but also a fundamental tool for inter-agent communication. This research presents explainability capabilities that have been introduced to the ASTRA agent programming language, which allow agents to store explanatory content and utilise configurable explanation generation algorithms, as preliminary steps in realising inter-agent explainability mechanisms in BDI multi-agent systems. As an illustrative example, it demonstrates agents engaged in learning using the implemented inter-agent explanations to deliberate over shared knowledge.
We propose a novel incentive-aware agent-based model of the E-Mini S P futures market with adaptive market maker agents and other non-adaptive background agents. Our agent-based model is qualitatively able to replicate the price dynamics that unfolded during the May 2010 flash crash. Circuit breakers (trading pauses) were one of the key regulatory measures that were introduced to avoid flash crashes, reduce volatility and improve resilience during times of market turmoil. However, several circuit breakers critics posit the existence of a ‘magnet effect’, where market volatility increases in the vicinity of a circuit breaker price level, thereby undermining the effectiveness of the trading pause. We analyse the adaptive behaviour of market maker agents during a flash crash using our proposed agent-based model to investigate the existence of ‘magnet effect’ under different market conditions, assuming these agents act rationally to maximise expected surplus. We find that the ‘magnet effect’ exists in high volatility market conditions driven by conservative inventory management and more extreme price skewing by market maker agents. This study aims to bridge the existing research gap by providing a framework for analysing adaptive response of agents to regulatory interventions during flash crashes and is a foundational step towards designing safer financial markets.
