Chester Ismay, a data science educator and AI consultant, built a sports-concierge agent to decide which games to watch each week. He follows several leagues — including the WNBA, NFL, NBA, and the Premier League — and the agent compiles a weekly summary based on his preferences and current schedules, then sends it to his phone.
System components
-
Preferences: Ismay began with a structured preferences file that lists the teams he follows and explains why he follows them. He also built a simple web interface to edit those preferences so he does not have to modify the underlying JSON directly.
-
Access to schedule data: His existing schedule viewers pull game schedules from sources such as ESPN and store the files in GitHub repositories. A read-schedules tool, run in Node, fetches the latest schedule files and merges them with Ismay’s preferences so the agent has the necessary information without searching for each game individually.
-
Agent policy: Much of the walkthrough focused on CONCIERGE.md, the policy file that defines the agent’s role. The policy states the goal, identifies data sources, and sets rules for choosing which games to recommend. It also explains how to treat completed tournaments and duplicate matchups, and specifies expected output and delivery format. When results are off, Ismay returns to this file to add detail or adjust instructions to better match the project goals.
-
Limited permissions: Ismay used Claude Code to read files, apply the policy, and generate the weekly recommendations. Permissions were configured so the agent only accesses the files and tools required for the task.
Scheduling delivery and validation
Ismay uses launchd on his Mac to run the concierge every Wednesday, and ntfy to send the generated summary to his phone. He validates the recommendations against the underlying schedules and uses tests plus multiple data sources to catch errors.
A notable implementation detail: time zone handling. The system could default to UTC and cause games to appear on the wrong day, so Ismay added explicit time zone instructions.
What’s next
The next episode of the series will feature AI engineer Sajal Sharma, who will give an agent its own computer in the cloud and demonstrate services and tools such as E2B and Scrapybara. He will show how sandboxing allows an agent to install packages, run code, drive a browser, and control a remote desktop.
Where to follow
The series "Zero to Agent in 30 Minutes" is available on Radar and in podcast form on YouTube, Spotify, Apple, and other platforms. O’Reilly members can watch live.
A closing note from the producers: if your work involves cybersecurity in any way, they invite you to answer an 11-question survey to inform a report.
