Intelligent agents could reshape pathology workflows

A new review published July 7, 2026, argues that AI agents may move pathology from single-image predictions to step-by-step diagnostic workflows that mirror how pathologists actually work. The authors say the approach could improve slide review, differential diagnosis and reporting, but only after rigorous testing proves the systems are reliable, traceable and safe.

Why it matters: - Intelligent agents could help pathologists review large digital slides, weigh evidence across multiple tests and write more consistent reports. - The approach matters most in complex oncology, rare diseases and settings with limited specialist expertise. - The review frames agent-based pathology as a possible shift from one-shot AI predictions to a more realistic diagnostic workflow.

What happened: - Researchers from the Department of Pathology at Peking Union Medical College Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College published a review online on July 7, 2026. - The article appeared in the Medical Journal of Peking Union Medical College Hospital. - The review, DOI: 10.12290/xhyxzz.2026-0402, examines how intelligent agents could support pathology from slide overview to diagnosis and report writing. - The source article is available here.

The details: - The review says computational pathology has advanced from narrow machine-learning tools to foundation models trained across organs, diseases and tasks. - Most current systems still analyze a slide or image in a single pass and return a classification, score or short description. - The authors argue that design misses how pathologists work in practice: they scan large tissue areas, change magnification, compare slides, combine morphology with immunohistochemistry and molecular tests, and weigh clinical information. - The review says intelligent agents could organize pathology into three linked stages: low-magnification slide overview, diagnostic reasoning and report generation. - The proposed systems would use large language models, vision-language models and specialized computational tools. - At the overview stage, navigation agents would decide where to move, when to zoom and whether enough information has been gathered. - Conventional slide analysis often uses whole-slide image patching and multiple instance learning, but those methods rarely model the search strategy itself. - At the diagnostic stage, visual models would extract morphological features while large language models coordinate reasoning. - Chain-of-thought methods could organize observed features, supporting evidence and competing diagnoses. - Tool use could let the agent call image-analysis, statistical or multimodal modules when needed. - Retrieval-augmented generation could bring in guidelines, textbooks or annotated cases, especially for rare or atypical findings. - For reporting, emerging systems try to combine information from several slides instead of generating text from one image. - Memory mechanisms could preserve previously observed features or retrieve similar historical cases. - Dynamic adaptation could incorporate pathologist feedback. - The review says incorrect memories, model drift, uncertain evidence tracing, privacy risks and inconsistent outputs remain major barriers.

Between the lines: - The review presents intelligent agents as a coordination layer, not just a better classifier. - That matters because pathology is iterative and evidence-based, and current AI pipelines often flatten that process into a single prediction. - The authors emphasize that clinical usefulness will depend on transparent reasoning, traceable evidence, controlled updating and evaluation with practicing pathologists. - Most existing studies still focus on visual question answering or restricted diagnostic tasks rather than end-to-end care. - The paper suggests the field is moving toward systems that can know what evidence is missing, choose the right tool and express uncertainty instead of forcing a conclusion.

What’s next: - Future work needs to integrate modules across all stages of the workflow. - The authors call for systematic trials to test stability, reproducibility, data security, auditability and clinical benefit. - Pathology agents will need to prove they can handle real diagnostic workflows before they become routine tools. - The review says dependable performance across full clinical use has not yet been demonstrated.