Producing large-scale political datasets demands extracting structured facts from unstructured sources, traditionally relying on expensive human experts and resisting at-scale automation. This paper develops and evaluates large language model (LLM)-based solutions to this bottleneck, focusing on elite biographies, one consequential class of political facts. We propose a two-stage ``Synthesis--Coding'' framework: LLM agents first search, filter, and curate evidence from heterogeneous web sources, then map curated inputs into structured records. We validate the framework across Chinese, American, and OECD political elites, benchmarking performance against human baselines using multiple state-of-the-art LLMs. We find that LLM coders match or exceed human experts when given curated inputs, and that agentic synthesis substantially outperforms human collective curation (Wikipedia) in open-web environments. We further identify a systematic bias: directly coding from long, multilingual corpora degrades extraction quality, and demonstrate that the synthesis stage mitigates this bias by compressing evidence into signal-dense representations.