AgenticDataBench benchmarks LLM data agents across 15 domains and 5 real-world B2B fintech workflows

Let's Data Science·medium signal

Researchers released AgenticDataBench, a benchmark for LLM-based data agents spanning 15 domains and five real-world B2B fintech use cases, evaluating realistic data-science workflows with fine-grained labels. A public GitHub testbed and Hugging Face dataset make the tasks reproducible and inspectable. It targets a gap in agent evaluation — end-to-end data workflows rather than isolated tool calls — giving builders a concrete yardstick for data-agent reliability.