She’s learning to drive in Colorado, and intersections are her weak spot: who goes first, and when can she turn? So I built her a game with AI agents, Can I Go Yet?

The first step was the rulebook every scene would be built on. The agent wrote it down, and one line said: “Right on red is allowed after a full stop.”

My reply to the list: “Awesome! Sounds great.”

That line is true at a red light. At a red arrow, Colorado doesn’t allow the turn at all, right turns included. A game exists to turn rules into reflexes, so she would have practised the wrong one until it felt automatic.

It never got that far. Before anything was built, a second agent got a single job: check every rule against the Colorado Driver Handbook and the state statutes. 12 minutes later it came back with the red arrow flagged.

Same AI, different job. The first agent wrote the rules from memory while designing a game. The second had nothing to defend and the law open in front of it.

When AI drafts something in your firm, who checks it against the source?