Beyond Scale: Building Truly Intelligent Multi-Agent Systems

Author: Denis Avetisyan

The pursuit of artificial intelligence must move beyond simply scaling up single-agent models to address the complexities of coordinated, adaptive behavior in multi-agent environments.

This review explores the challenges and emerging research directions for developing foundation models with native multi-agent intelligence, focusing on data, training, evaluation, and safety.

While foundation models increasingly serve as the core intelligence of AI agents, simply scaling single-agent capabilities does not guarantee robust performance in collaborative settings. This limitation motivates our work, ‘Towards Foundation Models with Native Multi-Agent Intelligence’, which argues for a dedicated focus on endowing these models with inherent multi-agent reasoning abilities-specifically understanding, planning, communication, and adaptation. Empirical evidence across 41 large language models demonstrates that strong single-agent performance is insufficient for achieving reliable multi-agent coordination, necessitating new research directions in dataset construction, evaluation metrics, training paradigms, and safety protocols. How can we best design and train foundation models to not only act intelligently, but also to collaborate intelligently?

Deconstructing the Monolith: The Rise of Collective Intelligence

The impressive capabilities of current foundation models, while groundbreaking for specific, isolated tasks, often fall short when addressing the intricate demands of real-world challenges. These problems are rarely solved by a single entity; instead, they necessitate a collaborative approach, mirroring the complex interactions observed in natural systems. Consider logistical operations, scientific discovery, or even social dynamics – all rely on the coordinated efforts of multiple actors. Consequently, the field is shifting towards multi-agent intelligence, recognizing that true problem-solving power emerges not from individual brilliance, but from the synergistic interplay of diverse perspectives and coordinated actions. This move acknowledges that intelligence isn’t simply about performing tasks, but about navigating and shaping complex environments through interaction and collaboration.

Existing artificial intelligence systems, often built around single, powerful agents, frequently falter when confronted with the unpredictable nature of real-world challenges. These systems, while adept at specific, pre-defined tasks, demonstrate limited capacity for nuanced reasoning – the ability to interpret context, handle ambiguity, and adapt to unforeseen circumstances. A critical shortcoming arises in dynamic scenarios, where conditions are constantly evolving and require ongoing assessment and adjustment. The rigidity of single-agent approaches prevents effective responses to these changes, emphasizing the necessity for systems capable of interaction, collaboration, and shared understanding to navigate complexity and achieve robust problem-solving.

The inherent limitations of single, monolithic artificial intelligence systems are driving research towards multi-agent architectures. Current AI frequently falters when faced with the ambiguities and shifting demands of real-world challenges, a consequence of operating in isolation. Consequently, a new generation of systems is being designed not as singular entities, but as collectives capable of coordinated action and mutual understanding. These systems require robust communication protocols, shared knowledge representations, and mechanisms for resolving conflicts or negotiating objectives. The development of such coordinated intelligence promises to unlock solutions for complex problems-from robotic swarms navigating dynamic environments to distributed networks optimizing resource allocation-by leveraging the combined strengths of multiple, specialized agents working in concert.

Mirroring the Social World: Understanding Agency in the Collective

Multi-Agent Understanding (MAU) is a critical capability for effective multi-agent systems, and fundamentally includes the capacity for Theory of Mind (ToM). ToM, in this context, refers to an agent’s ability to infer the beliefs, intentions, and emotional states of other agents within its environment. This inference is not simply recognizing states, but attributing them – understanding that other agents may hold beliefs different from its own, and that these beliefs will drive their actions. A robust MAU, therefore, enables agents to predict the behavior of others, coordinate effectively, and resolve conflicts, moving beyond simple reactive responses to more strategic and collaborative interactions. Without this capacity, agents are limited in their ability to function effectively in complex, dynamic multi-agent scenarios.

ToMBench and EmoBench serve as standardized evaluation frameworks for assessing an agent’s capacity for Multi-Agent Understanding. ToMBench specifically tests Theory of Mind – the ability to attribute beliefs, desires, and intentions to other agents – through tasks requiring inference of another agent’s knowledge or false beliefs. EmoBench, conversely, focuses on evaluating an agent’s capability to recognize and reason about emotional states, presenting scenarios that necessitate identifying and responding appropriately to the emotions of other agents. Both benchmarks utilize curated datasets and defined metrics, allowing for quantitative comparison of different agent architectures and training methodologies in the domain of social intelligence and collaborative problem-solving.

Evaluations of large language model scaling demonstrate differing performance improvements depending on the task domain. Specifically, increasing model size from Qwen-1 to Qwen-3 (both 8B parameter models) resulted in a significant increase in single-agent task accuracy, moving from a score of 0.23 to 0.64. However, performance gains on tasks specifically designed to evaluate Multi-Agent Understanding capabilities were considerably smaller, increasing from 0.44 to 0.55 over the same model scaling. This discrepancy suggests that improvements in general language modeling ability do not directly translate to equivalent progress in reasoning about the beliefs, intentions, or emotional states of other agents.

Orchestrating Action: Planning and Adaptation in the Swarm

Multi-Agent Planning (MAP) addresses scenarios where multiple agents must collaborate to achieve a shared objective, requiring the coordination of individual actions and consideration of inter-agent dependencies. Performance in MAP is frequently assessed using benchmarks such as CoordinationQA and Coordination Games, which present agents with tasks demanding joint reasoning and strategic interaction. These benchmarks evaluate an agent’s capacity to not only formulate its own plan but also to anticipate the actions of others and adjust its behavior accordingly. Successful MAP is crucial in applications ranging from robotics and autonomous driving to resource management and collaborative problem-solving, where decentralized decision-making is necessary to accomplish complex goals.

Efficient communication is a fundamental requirement for successful coordinated action between multiple agents. This necessitates the transmission of concise and precise information, minimizing ambiguity and reducing the potential for misinterpretation. The effectiveness of communication is not solely dependent on bandwidth or transmission speed, but critically on the semantic clarity of the exchanged messages. Agents must be capable of both accurately encoding their intentions and reliably decoding the intentions of others, enabling synchronized actions towards shared objectives. Failures in communication, even minor ones, can propagate through the system, leading to decreased overall performance and potentially preventing the successful completion of tasks requiring coordinated effort.

Evaluations of large language models Qwen (8B and subsequent iterations) and LLaMA (72B and subsequent iterations) on multi-agent planning benchmarks demonstrate a lack of correlation between single-agent performance improvements and gains in coordinated task completion. Specifically, accuracy metrics for these models, when tested on tasks requiring multi-agent coordination – such as those found in CoordinationQA and Coordination Games – have plateaued, remaining within the 0.2 to 0.35 range despite increases in model scale from Qwen-1 to Qwen-2.5 and LLaMA-2 to LLaMA-3.1. This suggests that enhancing a model’s general language understanding and reasoning abilities is insufficient for achieving effective multi-agent planning; dedicated mechanisms for coordination and communication are likely necessary.

Deconstructing the Bottleneck: Democratizing Multi-Agent Research

The accessibility of open-weight large language models (LLMs), specifically families like Qwen and LLaMA, is fundamentally altering the landscape of multi-agent system research. Previously, resource constraints – particularly the high costs associated with training and deploying proprietary LLMs – limited participation to well-funded institutions. These open-weight models, released with permissive licensing, allow researchers with limited computational resources to experiment with, fine-tune, and build upon state-of-the-art language capabilities. This democratization fosters broader innovation, accelerates the pace of discovery in multi-agent intelligence, and enables a more diverse range of perspectives and contributions to the field. The availability of pre-trained weights significantly reduces the barrier to entry, allowing researchers to focus on novel algorithmic development and system architectures rather than foundational model training.

vLLM is a fast and easy-to-use library for LLM inference and serving, designed to accelerate the development and evaluation of multi-agent systems. It utilizes PagedAttention, a novel attention algorithm that improves throughput by up to 24x compared to traditional attention mechanisms. This optimization is achieved through efficient memory management and reduced redundant computation during inference. vLLM supports continuous batching of incoming requests, further enhancing throughput and reducing latency. The framework’s streamlined deployment process, coupled with integrated evaluation tools, enables researchers and developers to iterate more rapidly on multi-agent system designs and assess performance metrics effectively. It is compatible with various model weights and supports distributed inference across multiple GPUs.

Population-Based Training (PBT) is an optimization strategy for machine learning agents that addresses challenges in dynamic and complex environments. Unlike traditional methods that optimize individual agents with fixed hyperparameters, PBT maintains a population of agents, each with its own set of hyperparameters. Throughout training, agents are periodically evaluated, and the best-performing agents “reproduce” by creating new agents with slightly perturbed hyperparameters, while poorly performing agents are replaced. This allows the population to collectively explore the hyperparameter space and adapt to changing conditions, leading to improved robustness and generalization. The process involves both exploitation – leveraging successful strategies – and exploration – discovering new, potentially superior approaches. PBT has demonstrated effectiveness in training agents for tasks requiring long-term adaptation and resilience to unforeseen circumstances, exceeding the performance of individually-trained agents in several multi-agent environments.

Beyond Scaling: The Emergence of Collective Intelligence

Recent advancements in foundation models have unlocked a new tier of problem-solving through multi-agent capabilities, exemplified by systems like OpenAI Operator and Kimi K2. These models don’t operate as singular entities, but rather orchestrate a collaborative network of specialized agents to dissect and resolve complex tasks. This distributed approach allows for a more nuanced and adaptable response than traditional single-agent models, as each agent can focus on a specific sub-problem or aspect of the overall challenge. Consequently, these multi-agent systems are demonstrating notable improvements in task completion across diverse benchmarks, suggesting a paradigm shift in how artificial intelligence tackles intricate problems by leveraging the power of collective intelligence.

Rigorous evaluation of foundation models equipped with multi-agent capabilities consistently demonstrates performance gains when measured against established single-agent benchmarks. Across diverse datasets – including the mathematical reasoning of MATH-500, the comprehensive knowledge assessment of MMLU-Pro, the code generation challenge of HumanEval, and the question-answering task of GPQA – these systems showcase improved accuracy and efficiency. These results aren’t merely incremental; they suggest that coordinating intelligence across multiple agents unlocks capabilities that simply scaling a single, monolithic model cannot achieve, providing quantifiable evidence of the benefits inherent in this architectural approach.

Research indicates that simply increasing the capabilities of a single artificial intelligence agent does not necessarily translate to improved performance when multiple agents collaborate. While larger, more powerful single agents excel in individual tasks, the benefits do not automatically scale to multi-agent systems. This suggests that effective collaboration requires more than just powerful individual components; it necessitates the development of specifically designed architectures and algorithms focused on native multi-agent intelligence – systems built from the ground up to prioritize communication, coordination, and collective problem-solving. This finding underscores the importance of shifting research focus beyond solely scaling single-agent prowess and towards understanding the unique challenges and opportunities presented by genuinely collaborative artificial intelligence.

The pursuit of truly intelligent multi-agent systems, as outlined in the paper, necessitates a willingness to dismantle established assumptions about scaling single-agent capabilities. This echoes Brian Kernighan’s sentiment: “Debugging is twice as hard as writing the code in the first place. Therefore, if you write the code as cleverly as possible, you are, by definition, not smart enough to debug it.” The paper champions a similar approach – deliberately stressing foundation models within multi-agent scenarios to reveal limitations and drive innovation. Rather than solely focusing on increasing scale, the research advocates for a systematic ‘breaking’ of current paradigms through focused dataset construction and novel training methods, ultimately revealing the underlying vulnerabilities and paving the way for more robust agent intelligence. The core idea is to push boundaries and understand failure modes.

Beyond the Scaling Law

The pursuit of increasingly large single-agent models has yielded impressive feats of pattern completion, but this work highlights a fundamental limit: sophistication does not inherently translate to cooperation. The illusion of intelligence, so easily conjured by a well-parameterized predictor, dissolves when faced with genuine interactive complexity. The next breakthrough won’t be a larger model, but a deeper understanding of how agency itself emerges within a collective. This requires a shift in focus-from optimizing performance on static benchmarks to constructing datasets that actively challenge the coordination and adaptation capabilities of these systems.

The proposed research directions – dataset construction, training paradigms, and safety – represent not merely incremental improvements, but potential exploits of comprehension. True multi-agent intelligence demands a reversal of the usual training protocol. Rather than providing examples of what to do, the goal should be to create environments that force the agents to discover effective strategies through interaction. This necessitates a re-evaluation of existing evaluation metrics, moving beyond individual performance to assess the emergent properties of the collective.

Ultimately, the safety considerations outlined are less about preventing malicious behavior and more about acknowledging the inherent unpredictability of complex adaptive systems. The real risk isn’t that these agents will choose to act against human interests, but that their interactions will produce unforeseen consequences. This field isn’t building intelligence; it’s reverse-engineering a fundamental process. And any sufficiently complex system, given enough freedom, will inevitably reveal its own limitations – and perhaps, its own surprises.

Original article: https://arxiv.org/pdf/2512.08743.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/