Abstract
This study addresses the problem of inefficient task scheduling and limited fault tolerance in ad-hoc grid computing environments, where traditional systems rely on centralized control and stable infrastructure. To overcome this, a decentralized agent behavior was developed using the Java Agent DEvelopment Framework, enabling autonomous task redistribution among heterogeneous Worker agents. The proposed Scheduler–Worker architecture allows dynamic coordination and failure recovery without centralized orchestration. Experiments on five devices show that increasing the number of agents from one to three reduces total execution time by 1.98–3.25, while the best performance is achieved with four agents, providing a 2.99 speedup for 100 tasks. However, using six agents on fewer devices reduces efficiency to 2.34 due to resource contention and communication overhead. The study is limited by a single network topology and a small-scale testbed. Nevertheless, the results demonstrate the practical potential of agent-based decentralized scheduling for resilient distributed machine learning systems.