DEPENDENCY MANAGEMENT IN SOFTWARE DEVELOPMENT PROJECTS USING A MULTI-AGENT APPROACH AND REINFORCEMENT LEARNING

Abstract

This paper addresses the problem of dependency management in large-scale software development projects, which are characterized by high complexity, dynamic requirements, and uncertainty. A multi-agent system model is proposed that integrates artificial intelligence methods and reinforcement learning to enable adaptive management of task dependencies. The architecture consists of three types of specialized agents: task agents that monitor task status and parameters; resource agents that optimize the allocation of human and technical resources; and risk agents that predict potential delays and conflicts. The central element is the coordinator agent, which, based on information from other agents, rebuilds the dependency graph, adjusts task priorities, and generates new execution scenarios. Optimization is achieved using the Proximal Policy Optimization (PPO) algorithm, allowing the agent to learn effective management strategies. Experimental evaluation was conducted on a test project in Jira, comparing the performance of a rule-based algorithm with that of the RL‑based coordinator agent. Evaluation metrics included the number of resource conflicts, average task risk, and critical path length. Results demonstrated that the agent can eliminate conflicts, shorten the critical path, and adaptively improve dependency management strategies. The key novelty lies in combining multi-agent architecture with reinforcement learning, which provides not only automation of routine decisions but also adaptive learning of optimal actions. At the same time, the model has limitations related to the quality of Jira data and the challenges of scaling to very large projects. The study contributes an experimental platform for further research on multi-agent systems in project management, highlighting their potential to enhance planning efficiency, reduce risks, and improve decision-making in dynamic software development environments