Seven AI Agents Turn a CSV Into a News Story

Researchers from Oxford and Stanford have built a system that turns a raw spreadsheet into a finished, interactive news article — and built it as a skill running on Claude Code. Called Data2Story, it coordinates seven AI agents that pass a CSV file down an assembly line, each handling one part of the job, and outputs a published web page where 93% of every visible statement traces back to its source: the data, the code, or an external link.

The seven roles read like a small newsroom. A Detective searches the web for context. An Analyst runs actual code to compute figures rather than estimating them. An Editor picks which findings drive the story. A Designer chooses the right format — a map for geography, a chart for trends. A Programmer builds the HTML, an Auditor checks the layout for errors, and an Inspector links every claim back to where it came from. Under the hood it runs on Claude Opus 4.7, pulling in outside models for images and media.

Readers liked it — with an asterisk

In a study of 53 readers comparing the agent's output against human-written originals across 18 datasets, 74% preferred the agent's article, 25% the human version, and 2% called it a tie. The agent won all five evaluation categories, with its biggest edge on transparency. The reason is that traceability number: 93% of claims verifiable, against a 25% baseline for human articles, where most statements simply aren't linked to a checkable source.

But the researchers are careful about what that means. Traceable is not the same as correct — it means you can follow a claim to its origin, not that the origin is right. And against The Pudding's design-heavy, handcrafted long-form pieces, the agent's lead vanished into a statistical tie. The system still loses on three fronts: editorial judgment about why something happened, genuinely creative design, and dense, single-visual graphics that pack a lot into one chart.

Why it matters for you

Most people who work with data don't write articles, but they do turn numbers into something others have to read and trust — a report, a dashboard summary, a slide. The interesting part of Data2Story isn't that it writes; it's that it forces every number to carry a receipt. That habit — make the claim, then show exactly where it came from — is what separates a trustworthy summary from a confident-sounding one.

It's also a concrete look at where multi-agent systems actually help. A single model asked to 'write an article from this CSV' tends to blur facts and prose together. Splitting the work — one agent computes, another verifies, another just checks links — produces output you can audit. That structure is reusable far beyond journalism, for any task where being able to check the work matters more than speed.

My take: the 93%-versus-25% gap is the line worth remembering. It's a quiet indictment of how rarely human-written summaries link their claims to anything — and a reminder that the bar an AI tool has to clear is often lower than we assume, because the human baseline already cuts a lot of corners.