MANTA: Adaptive Architecture for Multi-Agent Collaboration Networks via Self-Evolving Topology

Background and Context

Large Language Model-driven multi-agent systems have emerged as a powerful paradigm for solving complex problems, leveraging task decomposition, role specialization, and real-time information exchange to achieve results that single models cannot. However, a critical limitation persists in current architectures: the communication topology is typically treated as a static design choice or optimized only in offline phases. This rigidity ignores the dynamic nature of task execution, where the optimal flow of information changes as the problem evolves.

Consequently, systems often struggle with novel or highly complex scenarios that require adaptive coordination strategies. To address this gap, researchers have introduced the MANTA (Multi-Agent Network Topology Adaptation) framework. MANTA represents a shift from passive execution to active adaptation, enabling the self-evolution of communication structures during the inference phase. By allowing the network to dynamically adjust its internal collaboration structure based on real-time task demands, MANTA aims to provide a more robust and flexible architectural paradigm for multi-agent collaboration.

Deep Analysis

MANTA implements a sophisticated two-stage adaptive mechanism that ensures topology updates are both flexible and strictly controlled. The process begins in the initialization phase, where the framework generates an initial communication topology based on prior structural experience. Rather than relying on random initialization, MANTA utilizes patterns accumulated from historical data to ensure the starting point possesses a reasonable foundation for collaboration. This conditional topology is tailored to the specific conditions of the incoming task, setting a baseline that is already informed by past successes and failures. This approach significantly reduces the search space for optimal configurations, allowing the system to converge on effective collaboration strategies more rapidly than systems starting from scratch.

During the deployment phase, MANTA continuously monitors the collaboration trajectories of the agents. It evaluates the efficiency and sufficiency of the current organizational structure in real-time. When the system detects that the existing topology is insufficient to support task progression, it triggers a bounded structural update mechanism. These updates are comprehensive, covering multiple dimensions of the network architecture. The framework can modify agent role definitions, adjust communication links between specific agents, change the execution order of tasks, control information visibility, and optimize verification paths. Crucially, all these structural changes are constrained to maintain consistency in task interfaces and keep agent budgets constant. This design ensures that performance gains are achieved solely through the reorganization of internal connections, without incurring additional computational resource costs. This engineering efficiency highlights the theoretical elegance of MANTA, demonstrating that significant performance improvements can be realized through architectural agility rather than brute-force resource allocation.

Industry Impact

The validation of MANTA was conducted across five diverse benchmarks covering information retrieval, tool use, planning, workflow execution, and mathematical reasoning. These tests were designed to simulate the varied and complex scenarios found in real-world applications, thereby testing the framework's generalizability. The results were compelling: MANTA achieved an average score of 74.0, surpassing the strongest single-agent and multi-agent baselines by 5.8 percentage points. Notably, MANTA obtained the best results on the PlanCraft benchmark, a test specifically designed to evaluate complex planning capabilities. This success underscores the framework's ability to handle tasks that require intricate coordination and dynamic strategy adjustment. Ablation studies further revealed that the runtime self-improvement mechanism contributed significantly to these performance gains, particularly in tasks where collaboration strategies needed to shift dynamically. Compared to models with fixed topologies, MANTA was more effective at leveraging the complementary capabilities of agents, reducing无效 communication, and enhancing the overall efficiency of information flow.

From an industry perspective, MANTA offers a new perspective on multi-agent system design. It demonstrates that dynamic adjustment of collaboration structures is a potent lever for performance improvement, independent of optimizing individual agent capabilities or prompt engineering. For the open-source community, MANTA provides a reproducible framework that encourages further exploration of topology-adaptive methods. In industrial applications, the ability to boost system performance without increasing hardware costs is particularly valuable for resource-constrained edge computing or large-scale distributed systems. As multi-agent systems are increasingly deployed in fields such as autonomous driving, financial trading, and intelligent customer service, the demand for high-reliability AI infrastructure grows. MANTA's approach to runtime self-evolution addresses this need by providing a mechanism for systems to adapt to changing environments and requirements in real-time, fostering greater resilience and efficiency.

Outlook

The introduction of MANTA marks a significant step toward more autonomous and adaptive multi-agent ecosystems. By proving that communication structures can evolve during inference, the framework opens new avenues for research into self-optimizing AI networks. The concept of runtime self-evolution is likely to inspire subsequent studies that extend adaptive mechanisms to other system levels, such as dynamic resource allocation and real-time policy adjustment.

As the complexity of AI-driven tasks continues to rise, the ability of systems to reconfigure themselves on the fly will become a key differentiator. MANTA's success suggests that the future of multi-agent collaboration lies not in rigid, pre-defined architectures, but in fluid, self-healing networks that can learn and adapt in real-time. This evolution from static to dynamic topologies will be crucial for building the next generation of reliable and efficient AI systems, enabling a substantive leap from individual intelligence to collective intelligence. The implications for industries reliant on complex, coordinated decision-making are profound, promising systems that are not only smarter but also more resilient and adaptable to the unpredictable nature of real-world operations.