Digital agents can now autonomously gather, organize, reason over information, and employ tools to accomplish complex tasks. Forecasting stresses these abilities: it demands integration of incomplete evidence, inference of causal drivers, and delivery of calibrated probabilistic judgments. Existing benchmarks are either live—providing feedback only after events resolve and thus tying evaluation cycles to the calendar—or fixed‑corpus, which forces research to stay within pre‑collected documents. To address this gap we introduce FABench, a scalable backtesting platform for forecasting agents. An automated pipeline discovers events, formulates questions, and resolves outcomes for any past forecast date, allowing the benchmark to stay ahead of model training cutoffs; the first release offers 287 questions across nine domains. Agents freely research each query in an archive environment that blends an offline snapshot with live web search screened to prevent temporal leakage. Evaluating agents built on eight recent models shows that retrieval consistently improves Brier scores across all settings and lifts the top agents to the market consensus on the 54 Polymarket‑linked questions, with the best three scoring slightly above that benchmark. Trace analysis reveals recurring failures in how agents weigh and apply evidence, indicating ample room for improvement. We release the code, data, and archive environment as a foundation for community research on forecasting agents.\n\nReview