This paper presents Agentic Neural Network (ANN), a novel framework that effectively performs complex and high-dimensional tasks by leveraging multiple large language models (LLMs). ANN conceptualizes multi-agent collaboration as a hierarchical neural network structure, with each agent as a node and each layer as a collaborative “team” focused on a specific subtask. It operates through a two-stage optimization strategy: a forward stage that dynamically decomposes tasks into subtasks and uses appropriate aggregation methods to form a team of collaborative agents hierarchically, and a backward stage that mimics backpropagation to improve global and local collaboration through iterative feedback, and where agents evolve their own roles, prompts, and coordination. This neurosymbolic approach allows ANNs to generate new or specialized agent teams after training, significantly improving accuracy and adaptability. On four benchmark datasets, ANNs consistently outperform leading multi-agent baseline models under the same configuration. In conclusion, ANNs provide a scalable and data-driven multi-agent system framework that combines the collaborative capabilities of LLMs with the efficiency and flexibility of neural network principles. We plan to open-source the entire framework.
