7 AI Agents Write an Article from a CSV File

A team of researchers from Oxford and Stanford has developed a pipeline of seven AI agents that generates an interactive news story from a CSV file. Dubbed Data2Story, the system processes the input through a series of seven agents, each responsible for different aspects of story generation, before publishing the final story as a web page. In this web page, 93% of all visible assertions are backed by data, code, or external references.

These agents form a miniature news team. There’s a Detective who looks up information on the internet, an Analyst who actually executes code to calculate metrics instead of guessing them, an Editor who decides what insights to focus on, a Designer who selects appropriate visualizations (a map for geographic data, a chart for time series), a Programmer who writes the HTML, an Auditor who examines the visualizations for mistakes, and an Inspector who ensures that each assertion is properly attributed to its source. Data2Story leverages Claude Opus 4.7 underneath while using external models for media and images.

1Readers liked it — with an asterisk

When evaluated on 18 datasets with 53 participants, the agent was chosen over the human article in 74% of cases, tied in 2% of cases, and chose the human article in 25% of cases. The agent scored first in all five criteria, particularly excelling in transparency. This is because the agent generates stories where 93% of assertions can be traced to their sources compared to only 25% in human articles, where many statements aren’t backed by any source that can be checked.

However, the authors caution that traceability doesn’t imply accuracy. Instead, it only implies that assertions can be traced to their source. When compared to the visually oriented, human-designed longform articles of The Pudding, the advantage of the agent becomes negligible, statistically speaking. The system falls short in three areas. Firstly, it lacks the ability to make editorial judgements about causal relationships. Secondly, it can’t generate truly creative designs. Finally, it can’t produce complex visualizations that convey information in a single graphic.

2Why it matters for you

Many data professionals generate content from data but rarely do they write articles. Instead, they prepare reports, dashboard summaries, and slides. What makes Data2Story exciting is not its ability to generate stories but its insistence on accountability. Each number comes with a receipt. When you make an assertion, you must also show how you arrived at it. This is the difference between an accurate and a convincing summary.

Data2Story provides a glimpse into how multi-agent approaches can be useful. When we ask a single model to ‘generate an article based on this CSV’, we run the risk of confusing facts and prose. By splitting tasks between different agents — one for computation, one for verification, one for checking links — we get results that can be checked. This pattern can be generalized to other domains where accountability is important.

My thoughts: The key takeaway here is 93% vs 25%. It’s a subtle critique of how infrequently human-authored summaries are backed by sources and a reminder that sometimes the standard set by humans is low enough that the AI doesn’t need to be perfect to surpass it.