Autonomous multimodal agents enable transparent, spatiotemporal reconstruction of immune dynamics in pancreatic cancer progression

Pancreatic cancer progression is orchestrated by dynamic shifts in immune and stromal cellular ecosystems, yet the temporal and spatial principles governing these transitions remain poorly understood. Here, we present an agentic computational pathology framework that leverages large language models to orchestrate modular biomarker inference and spatiotemporal reasoning directly from routine H&E histology. Our approach, ROSIE (RObust in Silico Immunofluorescence), combines deep-learning-based multiplex inference with LLM-driven agent logic that emulates pathologist-level reasoning, enabling transparent and reproducible analysis of complex tissue microarchitectures. Applying this workflow to pancreatic intraepithelial neoplasia (PanIN) progression in KSC transgenic mice (n=24, ages 4-12 weeks), we generated 10.44 million single-cell profiles and identified a temporally ordered immune trajectory comprising three spatially distinct immune-stromal states: (1) early immune-surveillance niche: sharply bounded window of adaptive immune activation and antigen-presentation enrichment; (2) transitional mixed state: declining lymphoid activity, emerging exhaustion programs, and early EMT/angiogenesis signals; (3) stromal-dominant terminal state: fibroblast expansion, vascular remodeling, and immune silence. These findings establish pancreatic cancer progression as a temporally ordered sequence of immune activation, exhaustion, and stromal takeover. The agentic framework transcends static AI models by offering dynamic, tool-augmented reasoning that bridges high-dimensional tissue data with clinical interpretability--providing a scalable foundation for identifying therapeutic inflection points in early tumor evolution.