Abstract
We present a decentralized two-layer architecture for dynamic task assignment in multi-agent systems, designed to operate under partial observability, noisy feedback, and limited communication. The system consists of adaptive controllers that predict task parameters via recursive regression with forgetting and selectively broadcast tasks to a small subset of agents based on relevance and availability. To ensure consistency of task models across the network, we introduce a distributed optimization procedure that combines Simultaneous Perturbation Stochastic Approximation (SPSA) with consensus-based synchronization. The proposed approach enables scalable, online task allocation without centralized coordination. As a representative application, we evaluate the system on simulated workloads involving prompt-based tasks assigned to a diverse set of large language models (LLMs), demonstrating its robustness across varying noise levels, task dynamics, and input arrival patterns.