Patent data is a key resource for innovation research, yet a significant portion of historical records remains inaccessible in unstructured formats. This project addresses the case of approximately 2 million USPTO patent documents from 1950 to 1980, which remain largely inaccessible for large-scale analyses. We explore the use of Large Language Models (LLMs), deployed as AI Agents, to digitize and extract information from these documents. This includes applying techniques such as few-shot prompting, Chain of Thought reasoning, and the use of Reflection and ReAct-based agents. Preliminary findings highlight the potential of LLMs for natural language processing and reasoning tasks, as well as differences between AI Agent approaches. Once digitized, the data can be linked to existing datasets, enabling longitudinal studies of innovation in the US across centuries.