A Multi-Agent Workflow Orchestration System for Complex Long-Cycle Tasks

Abstract

A dynamic workflow orchestration system with the control plane separated from the execution plane is proposed to address the problem that traditional hard-coded service orchestration cannot handle business process changes and service failures in long-cycle tasks. The system uses a Multi-Agent architecture, through the division of roles such as the orchestration engine Agent, task execution Agent, and user Agent, combined with directory service (DF) registration and QOS-driven Contract Network Protocol (CNP), to achieve dynamic service selection and delay binding for flexible activities. An access control model based on dynamic binding of roles and tasks (R-TBAC) is introduced to enable dynamic granting and revocation of permissions driven by task instances.