Building an AI Analyst, One Football Argument at a Time
O'Reilly, Monday, October 19th, 2026
A hands-on series builds a data-analysis agent on NFL play data, showing how scoping and visible filters affect answers.
Chester Ismay and Michelle Smith start an eight-part series building an analytics agent on a quarter-million NFL plays.
In the first episode they test the claim that teams must establish the run: the agent writes down what would need to be true in the data, uses a read-only tool for the counting, and follows a written policy that bars causal claims and out-of-scope questions such as betting.
A whole-game view showed run-heavy teams winning 77.7% of games, but restricting the analysis to first-half running in close games reversed the result.
The authors' lesson is that filters, row counts and code should be part of every agent answer.