Rules all the way down: how AI agents get overhobbled
You can’t run creative agents without rules. Without them, the agent goes off the rails. But a rulebook rots in a specific way, and we only noticed when it was already shaping every show we make.
Each rule starts as a sensible local fix. An episode goes wrong, someone writes a line to stop it happening again, and the line cites the incident. Nobody ever re-reads the whole set as one text. A hundred fixes later, you have contradictions nobody wrote on purpose. Our playbook said “restating is fine” in the same paragraph as “subtext stays subtext”. Our script reader asked “does a line say why this object matters?” and then closed with “never an explanation line”. Neither rule was wrong on its own.
Three things make this worse with agents than with people:
- Models follow rules literally and resolve conflicts arbitrarily. A human writer reads “never explain” next to “make it clear” and finds the sensible middle. A model picks one, and may pick the other next time. So the symptom isn’t a visible bug. It’s quality that wobbles from episode to episode.
- Taste gets mistaken for craft. “Subtext stays subtext” is excellent advice for a prestige drama. Written into the core rules, it quietly becomes law for every show, including ours, made for someone half-watching on a phone at midnight. It took a colleague’s outside eye ("stop being clever: TikTok brain is flash memory") to see it.
- Reviewers multiply a rule. One line in our viewer-judge’s rules runs at four stages per show. A single bad line reaches everything.
What seems to keep it in check:
1. Layers: core craft, house taste, the show’s bible. Most hobbling is a rule sitting in the wrong layer.
2. Every rule carries its why. Without the why, nobody can later tell whether the rule still applies.
3. Invariants, not procedures. “The payoff plays out” survives new situations. A list of trimming steps doesn’t.
4. Test rules against outcomes. We had blind readers score a dozen first episodes and compared the results with how far real viewers got into each one. That turns taste arguments into evidence.
5. Read the whole thing end to end now and then. Our audit took one agent about five minutes and found 18 rules sitting in the wrong layer. It doesn’t happen on its own, because each change only looks at its own neighbourhood.
The lesson isn’t “fewer rules”. It’s that a rulebook is a codebase. It needs layering, comments and the occasional full read, or it quietly starts writing your shows for you.