A method for retrospective adaptation of contextual ai-agent weights and autonomy thresholds in IT projects

Abstract

This paper addresses the scientific and applied problem of retrospectively refining collective decision parameters for specialized artificial intelligence agents in a dynamic IT project. The study aims to develop a method that compares local agent forecasts with actual outcomes after a managed episode, adjusts contextual influence weights, and consistently changes the autonomy thresholds applied to subsequent decisions. The method separates operational and retrospective loops. In the operational loop, weights depend on role fit, competence, data quality, and accumulated agent reliability within a specific context class. In the retrospective loop, an agent contribution is evaluated by the change in collective loss with and without that assessment rather than by agreement with the majority. Confidence and risk thresholds are adapted using smoothed frequencies of unsafe autonomous execution and unnecessary escalation. Candidate parameters are introduced only after validation and never alter explanations of completed decisions. Experimental validation used a reproducible simulation containing one thousand nine hundred and twenty episodes, six specialized agents, four alternatives, and four context classes. Controlled accuracy drift was introduced for one specialist in the middle portion of the sequence. The complete method achieved an alternative-selection accuracy of approximately seventy-two per cent and reduced mean normalized regret by nearly thirty-seven per cent compared with unweighted aggregation. At a similar autonomy rate, unsafe autonomous execution occurred less than half as often as with fixed contextual weights and thresholds. The scientific novelty lies in context-localized retrospective updating of agent weights and joint adaptation of autonomy thresholds with validation gates, parameter versioning, and a non-retroactivity requirement. The practical value is a reduced dependence of collective decisions on obsolete agent reputation and static automation rules while retaining explicit human control over high-risk cases.