A model of collective decision-making and a method for context-dependent coordination of artificial intelligence agents in IT project management

Abstract

This article examines collective adaptive management decisions by specialized artificial intelligence agents in IT projects. It aims to improve decision validity, consistency, and adaptability by refining a collective decision model and an agent coordination method. The model combines project state, active roles, feasible alternatives, local assessments, conflicts, and execution results in one adaptive cycle. Results are compared with forecasts and used to update project state, performance, trust, and influence weights. Conflict is defined as a contradiction exceeding the difference normally expected between roles in the current context. Version consistency, feasibility, traceability, and stability are enforced as invariants. The coordination method detects contradictions and dynamically changes roles and authorities according to project state, agent competence, conflict graph density, uncertainty, and reconfiguration costs. The algorithm creates a versioned snapshot, activates roles, forms and evaluates alternatives, selects a feasible configuration, allocates tasks and resources, determines autonomy, and updates parameters after execution. Hysteresis and a frozen horizon limit reassignments, while configuration and plan are introduced atomically. Testing involved eighteen developers, three subteams, one hundred and twenty user stories, and six sprints lasting two weeks. Confidence increased from zero point seven one to zero point eight nine; autonomous dispatching from zero point five eight to zero point eight one; and risk mitigation effectiveness from zero point four eight to zero point seven two. The resolved conflict rate was zero point nine one, the response stability index was zero point eight seven, and the false escalation rate was zero point one three. Results confirm overall effectiveness.