A Multi-Agent Generative AI Architecture for Automated Task Planning, Collaboration, and Intelligent Decision Support

Abstract

This study proposes a hierarchical multi-agent generative artificial intelligence architecture for decomposing complex user objectives, assigning subtasks to specialized language-model agents, coordinating inter-agent collaboration, validating intermediate evidence, and delivering explainable decision recommendations. The architecture combines a supervisory planner, capability-aware scheduler, domain experts, retrieval and tool agents, independent critic, and policy-governed decision aggregator. A directed acyclic task graph represents dependencies, while multi-objective assignment balances expected quality, execution latency, monetary cost, and risk. A confidence-gated revision mechanism controls unsupported claims and escalates uncertain decisions for human review. An illustrative, explicitly synthetic benchmark of 200 tasks is provided to demonstrate an evaluation protocol across four system variants; the numerical values are not empirical findings. The proposed methodology outlines reproducible measurements for task success, plan validity, groundedness, latency, collaboration overhead, and decision utility. The contribution is a unified design and testable hypothesis about the circumstances under which structured multi-agent coordination improves decision support compared with single-agent generation.