I vibe coded a tiny MUD-style world sim where LLMs control each character. It's basically a little toy sandbox where LLMs can play around. There's no real goal to this, I just thought that it would be fun, like a more advanced tamagochi.

One of the issues I encountered initially was that the LLMs were repeating a small set of actions and never trying some of the more experimental actions. With a bit of prompt tweaking I was able to get them to branch out a bit, but it still feels like there's a lot of room for improvement on that front. I still haven't figured out how to instill a creative spark for exploration through my prompting skills.

It has been quite exciting to see how quickly a few simple rules can lead to emergent storytelling. One of the actions I added was the ability for the agents to pray to the creator of their world (i.e. me) along with the ability for me to respond in a separate cycle. The first prayer I received was from an agent that decided to wade into a river and kneel, just to offer a moment in stillness. Imagining it is still making me smile.

Unfortunately, I don't have access to enough compute to run a bigger experiment, but I think it would be really interesting to create lots of seed worlds / codebases which exist in a loop. With the twist being that after each cycle the agents can all suggest changes to their world. This would've previously been quite difficult, but I think it could be viable with current agentic programming capabilities. I wonder what a world with different LLM distributions would look like after a few iterations. What kind of worlds would Gemini, Claude, Grok, or ChatGPT create? And what if they're all put in the same world, which ones become the dominant force?