A model of context-dependent coordination and a method for the adaptive selection of interaction mechanisms between AI agents in IT project management

Abstract

This article addresses the challenge of adapting the coordination of specialized artificial intelligence agents to the dynamic state of an IT project. It is demonstrated that a static composition of agents and a pre-defined mechanism of interaction do not simultaneously ensure the required quality and speed of decision-making when requirements, deadlines, resources, conflicts and risks change. For the first time, a model of context-dependent coordination of specialized AI agents has been developed, which forms the complete configuration of the team based on the assessed state of the project: active agents, roles, communication links, coordination mechanism, weights and level of autonomy. Unlike static architectures, the model filters out structurally and normatively unacceptable configurations, assesses expected performance in the current context, and takes into account the cost of reconfiguration. A method for the adaptive selection and switching of coordination mechanisms for AI agents has been developed for the first time. The method derives a coordination requirement profile, generates a set of admissible mechanisms and performs a robust ranking of these based on decision-making quality and speed, suitability for the level of conflict, communication costs, risk and the results of previous application in similar contexts. A switch is made only when the advantage of the new mechanism exceeds an adaptive threshold, taking into account uncertainty, the structural cost of transition and the level of coordination required. In the absence of a viable option, a safe human-controlled mechanism is used. Experimental testing was carried out on an IT project scenario involving eighteen developers, three sub-teams, one hundred and twenty user stories and six two-week sprints. Compared with the best fixed mechanism, the proposed approach increased the average decision quality from zero point eight two to zero point nine zero, reduced the delay from five point one to three point six minutes, reduced residual conflict from zero point two three to zero point one two, and reduced normalised communication costs from zero point six one to zero point four seven. Ablation of the historical evaluation and hysteresis confirmed the contribution of these components to the quality and stability of the selection.