AgenticNet: Rethinking Multi-agent System Architectures with LLM-based Networks

Abstract: Multi-agent systems (MAS) enhance the capabilities of single LLM agents by leveraging collaboration and specialization. However, existing designs often rely on ad-hoc coordination strategies, lacking a principled architecture that integrates reasoning, communication, and adaptation. This limitation makes it difficult to scale multi-agent systems in a way that is both effective and interpretable. To address this challenge, we take inspiration from the architecture of neural networks to rethink MAS design. We treat each LLM-based agent as a computational unit and organize them into layered structures, analogous to neurons and layers, which we call AgenticNet. In this framework, lower layers act as planners that decompose problems, intermediate layers function as executors that advance reasoning step by step, and upper layers serve as synthesizers that verify consistency and deliver final decisions. Information propagates forward through the layers, while adaptation is guided by layer-level prompt updaters and a global prompt supervisor that refine agent behavior based on task loss, serving as an analogue to backpropagation. We conduct extensive experiments on five benchmarks, including AIME24, GSM8K, MATH500, HumanEval, and MBPP. Across all tasks, AgenticNet consistently outperforms both single-agent baselines and existing multi-agent systems, demonstrating its effectiveness as a scalable architecture for multi-agent.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Tim_Georg_Johann_Rudner1

Submission Number: 7332