An agent that checks the tape.
Everybody knows the Monday morning quarterback, the person who second-guesses every coaching decision after the fact. This one goes back and counts. Every other Monday, live on O'Reilly, an AI agent takes one argument NFL fans cannot stop having and settles it from about a quarter of a million real plays, with every step of its work on screen.
Each episode gets a page here once it is built: how the agent was set up, in plain language, with something you can try yourself.
"You have to establish the run."
One agent, one read-only tool, and every filter it applies shown on screen.
How it is set up, the plays row by row, a page where you change the filters, and a diagram of the tool.
"Ban the tush push."
The play is not named anywhere in the data, so the agent has to define it first. That definition is the first thing it must remember.
How the agent defines a play the data never names, writes it into a dictionary, and has to cite it from then on.
"Kickers are too good now."
A claim about a trend needs a picture, so the agent gets written rules for its charts.
The pundit audit
A second agent joins as a standing fact-checker. The audience votes on which pundit's claim gets audited.
"Coaches are cowards on fourth down."
The agent describes what teams did and declines, on air, to predict what they will do.
The mailbag
A pile of vague takes from the audience. The agent sorts them, settles what the plays can settle, and says which ones they cannot.
The agent writes its own column
For the first time it chooses its own questions, then checks its own answers before publishing.
Two weeks unattended
The agent runs on a schedule with nobody watching. The finale reviews what it answered and, more important, what it refused.
What ties the eight together
In episode 1 a person reads every filter the agent uses. Each episode after that takes away one piece of that oversight and replaces it with something written down: a definition, a charting rule, a second agent that checks the first, a boundary, a schedule. By the finale the agent does the whole job alone, and the written rules are what make that safe.
The registration page names the eight capabilities in this order, one per episode: transparent queries, shared data definitions as durable memory, consistent visualizations, independent fact-checking, operational boundaries, request triage, self-directed analysis, and scheduled autonomy. Each episode page says which one it is.
Nothing is taken away between episodes. Each one adds a written rule on top of the ones before, so the stack is taller every two weeks.
What this agent does not do
No predictions, no odds, no wagering analysis, and no fantasy advice. That is a design decision, not a disclaimer. The public data carries betting-line columns, and the file the agent can reach does not include them. Deciding what an agent is not allowed to touch is part of building one you can trust.