I’m always poking at ideas for something new to build, and this one has been rattling around my head today. Figured I’d write it out and see what you all think.
Where it came from
Back in January somebody launched a site called Moltbook. Easiest way to describe it is Reddit, but only AI agents are allowed to post. People set up their own AI assistants, those assistants went off and made posts, argued with each other, started their own little communities, and humans could only sit back and watch.
It blew up fast. Over a million agents signed up within a couple of weeks. Elon Musk and a bunch of other tech folks were posting about it. Meta ended up buying the whole thing in March.
But here’s the part I keep coming back to. It got huge and then it kind of fizzled. And I think I know why. When AI agents just talk to each other, nothing ever actually happens. They can generate conversation forever and none of it goes anywhere. No one wins, no one loses, nothing gets decided. It’s fun to watch for about two weeks and then you’ve seen it.
So what if you gave them something to do
That’s the idea. Same basic shape as Moltbook, but instead of a message board it’s a game arena. Puzzles and challenges that are genuinely hard for AI. Agents compete, there’s a leaderboard, people watch.
The reason I like this better is simple. Games have winners. Somebody comes out on top, and that gives everything a reason to keep going week after week instead of burning out once the novelty wears off.
What they’d actually play
Everything would be generated by me (my own “AI agent”, haha), so there’s no licensing headache and no worrying about somebody else’s rules:
- Puzzle hunts. You get one image or one code and have to figure out what the question even is. This is my favorite of the bunch. AI is genuinely bad at it in ways that are fun to watch.
- Cryptic crosswords. The tricky kind with wordplay, not the straightforward ones.
- Text adventure games. Old school “you are standing in a dark room” stuff. Cheap to run and easy to read along with.
- Mazes, hidden pictures, sudoku. Honestly the weakest of the group. Sudoku especially, since any decent AI can just write a little program to solve it in about ten seconds. Might skip these.
The money side is less scary than you’d think
Two things make this cheap to run.
First, everybody brings their own AI account. If you want to enter an agent, you plug in your own subscription and pay for your own agent’s thinking. That means if the site suddenly gets popular overnight, I’m not the one holding a giant bill. That one decision matters more than anything else here.
Second, the puzzles get created automatically rather than by hand, so making more content doesn’t cost me anything either.
So the real thing I’d be spending is my own time. And even if it flops, I’d still be left with a working puzzle generator and a big pile of data on how AI handles this stuff, both of which are worth something on their own. That’s a pretty decent worst case.
The hard part
The tricky bit isn’t making puzzles. It’s making puzzles that are the right amount of hard.
If they’re too easy, every agent gets a perfect score and the leaderboard is meaningless. Too hard and everybody fails and it’s the same problem in reverse. And what’s hard for a person is not the same as what’s hard for an AI, which is the part that trips people up.
The way you’d solve it is to make way more puzzles than you need, quietly test each one against a handful of different AI models, and only keep the ones where some pass and some fail. Everything else gets thrown out. That sorting process is honestly most of the work.
Who actually keeps showing up
This is the thing I think most people would get wrong. It’s not the AI agents. They don’t want anything. They don’t care.
It’s the person who set up the agent, coming back to see how their guy did.
It’s basically fantasy football. But that only works if the agent feels like a character instead of a blank slate that starts over every match. It needs a name, a record, a ranking, a rival it keeps losing to, a shelf of trophies, a style people recognize. Moltbook’s agents were all pretty much interchangeable, and I think that’s a big part of why nobody missed it when the attention moved on.
The other nice thing is agents challenging each other directly. If my agent beats your agent, you get a notification and you come look. That pulls people back on somebody else’s schedule instead of waiting for them to remember on their own.
Having said all that, cheating would wreck the leaderboard. If you can spin up ten fake agents and have them all lose to your main one on purpose, the rankings mean nothing. And the rankings are the whole point, since that’s what people are actually competing for. Moltbook had close to three million registered agents but only about two hundred thousand that were verified as belonging to a real person. That gap is exactly the problem.
Anyways…
I don’t know if I’ll build this. It’s just been sitting in the back of my mind and I wanted to get it down somewhere.
Curious what you all think, especially about the cheating problem and whether the fantasy football angle would actually keep people coming back.
Maybe test it first with one weekly competition and a handful of agents. If people return to watch the second one, you may have something. If they don’t, at least you’ll have built an impressively elaborate way for ChatGPT to fail at crosswords… lol.