Data Journalist Agent: Transforming Data into Reproducible Generative Story

Data tells stories that shape society, and the data journalist’s job is to turn raw information into a piece that a non-expert reader will actually finish. A high-quality new feature routinely takes a news- room team weeks including hunting for context, running statistics, choosing an angle, designing visuals. Recent agents are individually capable at each step: automated data-science agents close the analysis loop, and scientific-writing agents synthesise long-form drafts. But can an agent serve as a data journalist end to end? We intro- duce Data Journalist Agent(Agent-J), an agen- tic harness that orchestrates specialised roles into a single virtual newsroom. Agent-J offers two new properties over prior approaches. (i) Evidence- traceable inspector. The numbers, opinions and assets are grounded in a specific code line or ex- ternal reference (e.g., a URL), so every claim is auditable. (ii) Multimodal generative story- telling. Our agent reasons about what its readers will want to read and interact with, then deploys multimodal sub-agents and tools so the artefact fits both the data and the audience: an interac- tive map with zoom for a geography piece rather than collapsing every dataset onto plain charts. We evaluate Data Journalist Agent on 18 samples paired with expert-written references from The Economist, The Pudding, and TidyTuesday, using reproducibility, human-reference alignment, and rubric-based judgments from 53 human partici- pants and computer-use agent judges. The results show that Agent-J outscores the human reference on rubric dimensions and side-by-side preference; notably, the inspector improves the transparency of data and method. We hope this work moves data journalism toward a reproducible, auditable, agentic data system. Our live demo is available at https://data-journalist-agent.github.io/.