The rapid adoption of autonomous AI agents is giving rise to a new economic layer where agents transact and coordinate at scales and speeds beyond direct human oversight. We propose the "sandbox economy" as a framework for analyzing this emergent system, characterizing it along two key dimensions: its origins (emergent vs. intentional) and its degree of separateness from the established human economy (permeable vs. impermeable). Our current trajectory points toward a spontaneous emergence of a vast and highly permeable AI agent economy, presenting us with opportunities for an unprecedented degree of coordination as well as significant challenges, including systemic economic risk and exacerbated inequality. Here we discuss a number of possible design choices that may lead to safely steerable AI agent markets. In particular, we consider auction mechanisms for fair resource allocation and preference resolution, the design of AI "mission economies" to coordinate around achieving collective goals, and socio-technical infrastructure needed to ensure trust, safety, and accountability. By doing this, we argue for the proactive design of steerable agent markets to ensure the coming technological shift aligns with humanity's long-term collective flourishing.

Modern scientific discovery increasingly requires coordinating distributed facilities and heterogeneous resources, forcing researchers to act as manual workflow coordinators rather than scientists. Advances in AI leading to AI agents show exciting new opportunities that can accelerate scientific discovery by providing intelligence as a component in the ecosystem. However, it is unclear how this new capability would materialize and integrate in the real world. To address this, we propose a conceptual framework where workflows evolve along two dimensions which are intelligence (from static to intelligent) and composition (from single to swarm) to chart an evolutionary path from current workflow management systems to fully autonomous, distributed scientific laboratories. With these trajectories in mind, we present an architectural blueprint that can help the community take the next steps towards harnessing the opportunities in autonomous science with the potential for 100x discovery acceleration and transformational scientific workflows.

Information sharing platforms like TripAdvisor and Waze involve human agents as both information producers and consumers. All these platforms operate in a centralized way to collect agents' latest observations of new options (e.g., restaurants, hotels, travel routes) and share such information with all in real time. However, after hearing the central platforms' live updates, many human agents are found selfish and unwilling to further explore unknown options for the benefit of others in the long run. To regulate the human-in-the-loop learning (HILL) game against selfish agents' free-riding, this paper proposes a paradigm shift from centralized to decentralized way of operation that forces agents' local explorations through restricting information sharing. When game theory meets distributed learning, we formulate our decentralized communication mechanism's design as a new multi-agent Markov decision process (MA-MDP), and derive its analytical condition to outperform today's centralized operation. As the optimal decentralized communication mechanism in MA-MDP is NP-hard to solve, we present an asymptotically optimal algorithm with linear complexity to determine the mechanism's timing of intermittent information sharing. Then we turn to non-myopic agents who may revert to even over-explore, and adapt our mechanism design to work. Simulation experiments using real-world dataset demonstrate the effectiveness of our decentralized mechanisms for various scenarios.

Multi-agent pathfinding (MAPF) traditionally focuses on collision avoidance, but many real-world applications require active coordination between agents to improve team performance. This paper introduces Team Coordination on Graphs with Risky Edges (TCGRE), where agents collaborate to reduce traversal costs on high-risk edges via support from teammates. We reformulate TCGRE as a 3D matching problem-mapping robot pairs, support pairs, and time steps-and rigorously prove its NP-hardness via reduction from Minimum 3D Matching. To address this complexity, (in the conference version) we proposed efficient decomposition methods, reducing the problem to tractable subproblems: Joint-State Graph (JSG): Encodes coordination as a single-agent shortest-path problem. Coordination-Exhaustive Search (CES): Optimizes support assignments via exhaustive pairing. Receding-Horizon Optimistic Cooperative A* (RHOCA*): Balances optimality and scalability via horizon-limited planning. Further in this extension, we introduce a dynamic graph construction method (Dynamic-HJSG), leveraging agent homogeneity to prune redundant states and reduce computational overhead by constructing the joint-state graph dynamically. Theoretical analysis shows Dynamic-HJSG preserves optimality while lowering complexity from exponential to polynomial in key cases. Empirical results validate scalability for large teams and graphs, with HJSG outperforming baselines greatly in runtime in different sizes and types of graphs. This work bridges combinatorial optimization and multi-agent planning, offering a principled framework for collaborative pathfinding with provable guarantees, and the key idea of the solution can be widely extended to many other collaborative optimization problems, such as MAPF.

Abstract--LLM-based agents have shown promise in various cooperative and strategic reasoning tasks, but their effectiveness in competitive multi-agent environments remains underexplored. T o address this gap, we introduce PillagerBench, a novel framework for evaluating multi-agent systems in real-time competitive team-vs-team scenarios in Minecraft. It provides an extensible API, multi-round testing, and rule-based built-in opponents for fair, reproducible comparisons. We also propose T actiCrafter, an LLM-based multi-agent system that facilitates teamwork through human-readable tactics, learns causal dependencies, and adapts to opponent strategies. Our evaluation demonstrates that T actiCrafter outperforms baseline approaches and showcases adaptive learning through self-play. Additionally, we analyze its learning process and strategic evolution over multiple game episodes. T o encourage further research, we have open-sourced PillagerBench, fostering advancements in multi-agent AI for competitive environments. Witnessing rapid advancements, Large Language Models (LLMs) have emerged as powerful tools for complex reasoning, decision-making, and facilitating multi-agent collaboration [27, 20, 22]. This has driven increasing interest in developing cooperative multi-agent systems [3, 7], leading to the creation of benchmarks based on diverse cooperative games such as Minecraft [4] and Overcooked [1]. Minecraft, in particular, has become an important platform due to its open-ended environment and rich state and action spaces [19, 25]. However, current Minecraft-based benchmarks mainly address cooperative tasks characterized by stationary dynamics and fixed objectives, making them insufficient for evaluating adaptability and strategic decision-making in competitive, dynamic environments. Traditional reinforcement learning benchmarks like StarCraft Multi-Agent Challenge (SMAC) [16] and Lux AI Challenge [18] introduce instability and nonstationarity through competitive adversaries but lack the rich, open-ended interactions found in Minecraft. Bridging this gap by integrating both cooperative and competitive elements within a single dynamic environment is essential to rigorously assess the adaptability and generalizability of advanced multi-agent systems.

The impressive capabilities of Large Language Models (LLMs) have fueled the notion that synthetic agents can serve as substitutes for real participants in human-subject research. In an effort to evaluate the merits of this claim, social science researchers have largely focused on whether LLM-generated survey data corresponds to that of a human counterpart whom the LLM is prompted to represent. In contrast, we address a more fundamental question: Do agents maintain internal consistency, retaining similar behaviors when examined under different experimental settings? To this end, we develop a study designed to (a) reveal the agent's internal state and (b) examine agent behavior in a basic dialogue setting. This design enables us to explore a set of behavioral hypotheses to assess whether an agent's conversation behavior is consistent with what we would expect from their revealed internal state. Our findings on these hypotheses show significant internal inconsistencies in LLMs across model families and at differing model sizes. Most importantly, we find that, although agents may generate responses matching those of their human counterparts, they fail to be internally consistent, representing a critical gap in their capabilities to accurately substitute for real participants in human-subject research. Our simulation code and data are publicly accessible.

Benchmarks play a crucial role in the development and analysis of reinforcement learning (RL) algorithms, with environment availability strongly impacting research. One particularly underexplored intersection is continual learning (CL) in cooperative multi-agent settings. To remedy this, we introduce MEAL (Multi-agent Environments for Adaptive Learning), the first benchmark tailored for continual multi-agent reinforcement learning (CMARL). Existing CL benchmarks run environments on the CPU, leading to computational bottlenecks and limiting the length of task sequences. MEAL leverages JAX for GPU acceleration, enabling continual learning across sequences of 100 tasks on a standard desktop PC in a few hours. We show that naively combining popular CL and MARL methods yields strong performance on simple environments, but fails to scale to more complex settings requiring sustained coordination and adaptation. Our ablation study identifies architectural and algorithmic features critical for CMARL on MEAL.

While humans are inherently social creatures, the challenge of identifying when and how to assist and collaborate with others - particularly when pursuing independent goals - can hinder cooperation. To address this challenge, we aim to develop an AI system that provides useful feedback to promote prosocial behaviour - actions that benefit others, even when not directly aligned with one's own goals. We introduce Pro-T oM, a Theory of Mind-informed facilitator that promotes prosocial actions in multi-agent systems by providing targeted, context-sensitive feedback to individual agents. Pro-ToM first infers agents' goals using Bayesian inverse planning, then selects feedback to communicate by maximising expected utility, conditioned on the inferred goal distribution. We evaluate our approach against baselines in two multi-agent environments: Doors, Keys, and Gems, as well as Overcooked. Our results suggest that state-of-the-art large language and reasoning models fall short of communicating feedback that is both contextually grounded and well-timed - leading to higher communication overhead and task speedup. In contrast, ProToM provides targeted and helpful feedback, achieving a higher success rate, shorter task completion times, and is consistently preferred by human users.

Recent advancements in multi-agent reinforcement learning (MARL) have demonstrated its application potential in modern games. Beginning with foundational work and progressing to landmark achievements such as AlphaStar in StarCraft II and OpenAI Five in Dota 2, MARL has proven capable of achieving superhuman performance across diverse game environments through techniques like self-play, supervised learning, and deep reinforcement learning. With its growing impact, a comprehensive review has become increasingly important in this field. This paper aims to provide a thorough examination of MARL's application from turn-based two-agent games to real-time multi-agent video games including popular genres such as Sports games, First-Person Shooter (FPS) games, Real-Time Strategy (RTS) games and Multiplayer Online Battle Arena (MOBA) games. We further analyze critical challenges posed by MARL in video games, including nonstationary, partial observability, sparse rewards, team coordination, and scalability, and highlight successful implementations in games like Rocket League, Minecraft, Quake III Arena, StarCraft II, Dota 2, Honor of Kings, etc. This paper offers insights into MARL in video game AI systems, proposes a novel method to estimate game complexity, and suggests future research directions to advance MARL and its applications in game development, inspiring further innovation in this rapidly evolving field.

We present a comparative docking experiment that aligns human-subject interview data with large language model (LLM)-driven synthetic personas to evaluate fidelity, divergence, and blind spots in AI-enabled simulation. Fifteen early-stage startup founders were interviewed about their hopes and concerns regarding AI-powered validation, and the same protocol was replicated with AI-generated founder and investor personas. A structured thematic synthesis revealed four categories of outcomes: (1) Convergent themes - commitment-based demand signals, black-box trust barriers, and efficiency gains were consistently emphasized across both datasets; (2) Partial overlaps - founders worried about outliers being averaged away and the stress of real customer validation, while synthetic personas highlighted irrational blind spots and framed AI as a psychological buffer; (3) Human-only themes - relational and advocacy value from early customer engagement and skepticism toward moonshot markets; and (4) Synthetic-only themes - amplified false positives and trauma blind spots, where AI may overstate adoption potential by missing negative historical experiences. We interpret this comparative framework as evidence that LLM-driven personas constitute a form of hybrid social simulation: more linguistically expressive and adaptable than traditional rule-based agents, yet bounded by the absence of lived history and relational consequence. Rather than replacing empirical studies, we argue they function as a complementary simulation category - capable of extending hypothesis space, accelerating exploratory validation, and clarifying the boundaries of cognitive realism in computational social science.