By Chester Ismay and Michelle Smith — their team found that the term "tush push" does not appear in their roughly 265,000 play descriptions covering 2021 through the current season. "Brotherly shove" was also absent, and "sneak" appeared only twice. That made defining the play the first task when their AI analyst tackled the "Ban the Tush Push" topic on the October 26 episode of The Monday Morning Agent.
The dictionary: the agent’s first durable memory
They recorded the chosen definition in a shared data dictionary so future runs must use the same criteria and explicitly state the definition’s limits. This durable entry prevents later queries from silently changing what a prior result meant.
Check whether the data can support each clause of your definition
The authors note that the dataset identifies the ball carrier but does not record whether teammates pushed him. It can indicate whether a snap came from shotgun, but it cannot reconstruct a team’s exact formation.
They examined five candidate definitions, ranging from a quarterback sneak with teammates pushing to any rush on third or fourth-and-short. Each choice changes which plays the analysis counts; the broadest definition would include running backs and thus depart from the contentious play most people mean.
Their recommendation: before letting an agent query a colloquial term, write the definition in plain language and go clause by clause to verify whether the available data supports it. A familiar label does not guarantee an operationalizable field.
Write down what your definition cannot measure
The team settled on: a quarterback run with one yard or less to go, from under center. The agent counts qualifying quarterback sneaks whether anyone pushes the runner or not. In other words, they are measuring a proxy for the tush push and cannot isolate how much a shove contributed to success.
Even "quarterback" required an operational decision. The tool’s fields do not include player positions, so the quarterback was identified as the player who threw the most passes for that team in that game — another assumption to check when interpreting results.
The dictionary entry should explain the play in plain language and list what the data cannot tell you. Otherwise readers could infer that a "tush push" result necessarily counts plays with a push. Keep the limitation with the definition so the agent can repeat it alongside results.
Give future runs the same definition
The dictionary stores one JSON file per definition with a fixed structure. An agent can add an entry but cannot change an existing one; a different definition requires a new name. That preserves the meaning attached to an earlier result.
Before each run, the script that launches the agent (the runner) saves a copy of the dictionary and adds its contents and date to the agent’s instructions. The agent first searches for an existing definition; if it finds one, it uses it. If not, it can write a new entry using the supported fields.
In the demonstration, the tool initially refused a "tush push" request because no entry existed; the agent then created a definition and queried the plays.
Verify the agent actually uses its memory
Writing "use the dictionary" in a policy is not enough. The system relies on instruction, check, and enforcement. The policy requires the agent to use and cite an entry. The tool accepts definitions from the dictionary rather than an improvised query-time definition. If the answer lacks the required citation, the runner retries once and then stops if it is still missing.
The team expanded the agent’s permissions so it can add a dictionary entry alongside querying the plays. Test whether your agent uses shared definitions and whether a reader can locate them in the response.
Read the conversion rate before deciding on a ban
Across five complete seasons, the defined play converted 78.3% of the time versus 70% for other short-yardage rushes — an 8.3 percentage-point advantage. The agent reported the advantage but did not treat that gap as sufficient evidence to justify a ban. Those figures describe conversion rates for their defined group of plays; they do not establish whether the NFL should ban a play, nor do they quantify how much pushing helped.
A viewer raised that very distinction during the discussion: whether a successful play should remain in the game requires human judgment. Conversion rates can inform that conversation but cannot resolve questions about safety or about what people want football to look like.
Next topic: "Kickers Are Too Good Now"
On November 9 the team will address "Kickers Are Too Good Now" and provide the agent with written visualization standards. An agent can generate a chart in seconds; the work now is to set rules for presenting trends and comparisons so readers know what the data actually supports.
O’Reilly members can sign up on the O’Reilly learning platform to join live on November 9, 8:00–8:45 am PT / 11:00–11:45 am ET. Everyone can follow along on Radar every other Monday, with a brief pause for the holidays. Watch the October 26 episode and the team will build on it in the next session.
