Сoordinated decision support for it projects using multi-agent artificial intelligence systems

Abstract

Relevance: The article examines adaptive resource allocation and task scheduling in information technology project management under stochastic uncertainty. Centralized planning methods often become unstable when resource availability, task priorities or project scope change during execution. Aim: The aim of the article is to develop and evaluate a coordinated decision support framework for dynamic information technology projects based on a risk-aware multi-agent architecture. Objectives: The study focuses on formalizing the interaction between Project Manager, Task, Resource and Risk agents; defining a utility-based task allocation mechanism; and evaluating the proposed approach under resource disruption and scope creep scenarios. Methods: The proposed system uses a decentralized architecture in which autonomous agents negotiate task assignments through a modified Contract Net Protocol and an auction-based coordination mechanism. The utility function combines competence matching, execution cost, accumulated fatigue and risk probability. The weighting coefficients are selected through an expert pairwise comparison procedure based on the Analytic Hierarchy Process. The experimental evaluation was performed using one thousand Monte Carlo simulation iterations on an adapted project scheduling dataset. Scientific novelty: The novelty of the study lies in combining risk-aware and fatigue-aware decision logic within multi-agent negotiation for information technology project management. Unlike approaches focused only on time or cost, the proposed framework explicitly considers human workload and assignment risk. Practical significance: The framework can support project managers in adaptive re-planning, workload balancing and proactive risk mitigation while preserving the interpretability of task allocation decisions. Results: The proposed system achieved a mean project duration of one hundred twenty-four days, compared with one hundred forty-two days for the static critical path baseline, one hundred thirty-five days for the greedy heuristic and one hundred twenty-eight days for the deep reinforcement learning-based scheduler. The standard deviation of project duration decreased to six point two days, and the resource utilization Gini coefficient decreased to zero point one eight. Conclusions: The results confirm that decentralized risk-aware multi-agent coordination improves project resilience, schedule predictability and workload balance in dynamic information technology project environments.