A team ran a Dungeons and Dragons game in their office, just as a social hour. Nobody offered them a choice between two paths. Somewhere in the session, instead of going into the woods to kill trolls for the reward, they built a business in the town, and it worked.
Jeff Randolph heard that story from a guest on his own podcast, and told it again in a TEDx talk at Avila University this year. Randolph has run D&D games since the late 1970s, and he calls what almost happened instead “railroading”: the game master already knows where the story goes, and walks the players there one step at a time. No risk, no surprise, just following directions with extra dice. That office team never got railroaded. They found a path nobody had planned for.
I’ve played D&D since I was a teenager, through plenty of video game versions, and these days I run campaigns with my own daughter and the rest of the family. A class gives a player a set of abilities and a role, not a personality. Two people playing rogues will still solve the same locked door differently, because a character sheet only sets up what they have to consider. The judgment stays theirs. But the role changes what they have to consider: a fighter and a wizard look at that same door and see two different problems, whether they want to or not.
I’ve caught myself saying some version of this out loud a few times lately: there’s all roads lead to home. Plenty of paths get you to the same place. Some are faster. None of them are the only right one.
You lose whatever they’d have found that you wouldn’t have thought of yourself.
Stifle the player who wants to build the business instead of killing the trolls, and you don’t just lose a fun choice. You lose whatever they’d have found that you wouldn’t have thought of yourself. The people at a table are there to enjoy solving a hard problem together, and the problem gets solved better because several different minds are on it instead of one.
None of that works without agreeing on the destination first. Randolph spends real time in his talk on what he calls session zero: the conversation before anyone rolls a single die, where the group settles what it actually wants out of the game, what everyone’s expecting, whether the whole table is signed up for the same adventure. Leaving the method open only works once that part is settled. Skip it, and freedom just looks like nobody agreeing on what they’re doing.
I had a version of that session-zero problem with my own team yesterday, about our website. I was done patching Elementor, almost daily, and told my team I didn’t care what replaced it, I just wasn’t fighting that page builder anymore. If it had been up to my own default, I’d have stuck with Elementor. It’s what my team and I already know. My developer Josh builds in Cornerstone instead, and I trusted his judgment over my own habit.
He had reasons that went past patch frequency. In Cornerstone, he builds a button once as a component and reuses it everywhere; change it in one place, and it updates across the whole site. He’d just finished testing one of the new AI page-builder tools against Elementor on a client project, and found the opposite problem: the page looked right at a glance, but every element underneath it was its own block of inline CSS, never touching the actual design system. It would hold up for exactly one page and fall apart the moment anything needed to change consistently.
I set the destination. I was done fighting a page builder that had turned into a daily chore. I didn’t set the path. Josh had already walked it, down at the level of code I don’t touch anymore, and he came back with a reason to go a specific direction I wouldn’t have known to ask for on my own.
I’m trying to make the same shift with the AI agents doing work for me now. Give an agent a destination and a set of rules worth protecting, then let it find its own path through them, and you get Josh’s answer: a real reason, grounded in how the page actually behaves once it ships. Hand it a destination and the exact steps to get there instead, and you get what Josh saw in that AI page builder: looks right at a glance, and falls apart the first time anything underneath it has to hold.
Good thoughts here, but I'm missing on the whole "steps" thing. How can you trust an agent to do the right thing if you can't audit it? and if you can audit it, can't you reverse engineer the steps? and if you can reverse engineer the steps, you don't need an agent. and if you know the steps, you don't need an agent. I'm not comfortable just handing off stuff to LLM's