Agentic artificial intelligence in radiology workflow: from image interpretation to report quality control

Abstract

Radiology AI has grown past the single-purpose detector. The newest systems, built around large language models (LLMs), chain together the steps a radiologist actually works through: triaging the worklist, retrieving prior imaging studies, processing images, drafting a structured report, checking for mistakes. When these modules are coordinated through an orchestration layer, they form a multi-agent system capable of managing multiple stages of the radiology workflow—software that handles stretches of the radiology pipeline with less human input at each stage. This review maps the evidence behind that shift, drawing on PubMed-indexed studies from 2023 to 2026. We begin with convolutional neural networks and foundation models, then follow the emergence of AI agents that observe, plan, and act inside clinical environments. We examine multi-agent architectures—specialized agents for image analysis, report drafting, error detection, and decision support—and ask what they actually deliver. So far, the data tell a consistent story: multi-agent cross-verification drives hallucination rates down; intelligent worklist triage cuts report turnaround time by up to 43.7% in some settings; GPT-4 catches 82.7% of report errors, matching human readers. But nearly all of this evidence comes from single-center, retrospective studies on curated data. Every systematic review reaches the same conclusion: the technology works in the lab and has not been proven in the clinic. We also discuss compound opacity—how layered agent interactions make decisions harder to trace—alongside poor reporting standards and a regulatory framework that was not designed for generative, continuously-adaptive software. Agentic AI, including multi-agent architectures, could reshape how radiology departments operate, but the field needs prospective, multi-center trials with standardized endpoints before claiming it already has.