Multi-level adaptive replanning method for IT projects by a collective of coordinated artificial intelligence agents

Abstract

This paper addresses the scientific and applied problem of adaptive IT-project replanning by a collective of specialized artificial intelligence agents. The study is motivated by the simultaneous occurrence of schedule, resource, cost and risk deviations and by the fact that locally rational agent proposals may contradict one another. A single-level procedure cannot distinguish situations in which changing one task is sufficient from situations that require dependency restructuring or complete replanning of the unfinished project scope. An adaptive IT-project planning model is proposed that integrates the task graph, successive plan versions, dynamic project state, observed events, deviations, their structural impact scope and a collective of specialized AI agents into a single “event – state – proposals – new plan – actual outcome” cycle. The collective coordination model represents contradictions through a typed conflict matrix and graph, determines contextual agent influence from functional responsibility, relevance, current confidence and historical reliability, and supports bounded concessions, repeated requests, temporary reassignment of coordination functions and escalation. The resulting multi-level method calculates an integral deviation signal, selects one of four intervention levels, generates feasible alternatives within a bounded change scope, rechecks their resource and risk feasibility, coordinates conflicting proposals and produces a collective decision under scenario uncertainty and controlled autonomy. The completed project scope remains unchanged, while plan instability is explicitly included in the decision criterion. An experimental study covering one thousand nine hundred and twenty test episodes compared the proposed method with a static plan, local-only adaptation, centralized selection and unweighted voting. Against the strongest baseline, unweighted voting, the proposed method reduced actual loss by twenty-three point one one percent, normalized regret by eighty-one point seven three percent and hard-constraint violations by sixty percent, while increasing decision quality by ten point one one percent. Ablation results confirmed the separate contributions of the dynamic-state model and the collective coordination mechanisms.