Foundation Models and AI Agents in Digital Pathology Imaging: A Systematic Review of Integration into the Clinical Workflow and Implementation Challenges.

The objective of this study is to systematically assess foundation model‑driven AI agents in digital pathology imaging, focusing on technical progress, clinical validation, and workflow integration barriers from an implementation science perspective, while clearly distinguishing foundation models as general‑purpose pretrained representations from AI agents as goal‑directed workflow systems. PubMed, Embase, Scopus, Web of Science, and IEEE Xplore were searched (2020-2026). Studies reporting foundation models or AI agents in pathology image analysis with clinical validation were included. Quality was assessed using QUADAS‑2, PROBAST, and CLAIM. Data extraction was enhanced to capture multi‑task adaptability, zero‑shot capabilities, and downstream evaluation protocols. Forty‑two studies were included. When evaluated on downstream tasks after fine‑tuning or linear probing, foundation models demonstrated excellent performance (AUC 0.85-0.97 across diagnostic tasks). However, only six studies (14.3%) conducted multicenter validation; prospective studies were < 5%. AI agents operated at L1-L2 levels, with workflow integration relying on HL7/FHIR and DICOM standards. Four barriers emerged: data heterogeneity, limited explainability, insufficient workflow‑level validation, and regulatory gaps. The review also identified that foundation models' lightweight adaptability-enabling strong performance across 5-15 downstream task families-remains under‑reported. Clinical adoption requires multicenter prospective studies, standardized reporting (CLAIM), interoperability standards (DICOM for Pathology), and adaptive regulatory pathways for AI agents in digital pathology.