Meta paper describes ABTest Agent, where validated code computes all reported numbers

Meta reports that its ABTest Agent lets the LLM choose and order analysis steps, while validated tools compute every number it reports.

A report template blocks output unless each table traces back to that code. Meta says plain result fetching made up 45% of requests in one study, and 52 of 97 skills checked whether results could be trusted. The source is a Meta paper summarized in a post by Rohan Paul on X.