Three Worlds, One Engine
Most AI demos trap you in a chat box with a single, omniscient assistant. Multi-Agent Land does something stranger.
You don't just chat with it. Instead, you drop a disruption into a living, digital ecosystem and watch a small troupe of AI characters argue, remember, judge, and reshape their world in response.
The Seed: "A village of stage props wakes up and argues about which fairy tale they belong to."
Type that in, and a forest grows around it:
- A narrator describes a mossy ticket booth opening in a tree root.
- A pocket actor announces it's collecting echoes to knit a ladder to the moon.
- A critic decides whether that ladder is worth keeping in the story.
You lean in and whisper, "A lantern starts whispering recipes," and the whole cast reacts. Nobody is following a script. The story is whatever the troupe makes of your nudge.
That’s the toy. Here is why it’s built the way it is.
You're watching a troupe, not querying a model
The unit of delight here isn't a clever answer—it's a cast. Each character is its own small AI specialist with a fixed personality and exactly one job:
- The Narrator: Sets the scene.
- The Dreamer: Wants impossible things.
- The Transformer: Alters whatever you throw into the mix.
- The Judge: Decides what actually becomes real.
They have moods. They remember what happened earlier. They talk past each other and occasionally catch fire.
Read their minds: every public line (front) carries a private thought (revealed) — generated by the model, shown to you, hidden from the other agents.
The Twist: You can read their minds
Every line a character says in public comes paired with a private thought it isn't sharing—the bluffer's quiet panic, the judge's actual reasoning, the gossip behind the smile.
[ Public Line ] —> "Please do not applaud yet; my shadow is still rehearsing."
|
[ Private Thought ] —> (I hope they didn't notice I forgot my next three lines...)
A toggle flips the cards over so you can see both at once. The drama you watch on the surface has a second, hidden layer underneath, generated live by the models. The AI isn't just decorating the experience; the AI is the experience.
Read their minds: Every public line (front) carries a private thought (revealed)—generated by the model, shown to you, but completely hidden from the other agents.
Three worlds are three files, not three apps
You'd be forgiven for thinking a collaborative storytelling game, a murder-mystery solved by a swarm, and a tense game of twenty questions are three different products. They're not.
They are the same engine wearing different costumes—and the costume is a small text file, not a new codebase.
| World Type | The Goal | The Mechanic |
|---|---|---|
| Collaborative Wood | Pure exploration | No winner; focus on growing strange, organic narratives. |
| Murder Mystery | Convergence | Agents debate clues until they crown the best theory. |
| Guessing Game | Secrecy | The engine keeps a secret and tracks who cracks it first. |
Every world is a simple configuration. Swap the cast, swap the goal, and the same machinery puts on a completely different show.
Today, we have eight worlds running on one engine. Adding the ninth means writing a text file, not shipping new code.
(How that's possible—agents that never call each other, only post to a shared, event-sourced log—is the subject of Part 3. For now: it really is just config.)
What we're actually going for
When we make a change to this project, we ask whether it makes the thing more delightful, more clearly AI-load-bearing, more original, or more polished. Those four words are the whole compass:
- Delightful — the bar is "would you show a friend?" The wood should feel alive, lead with whimsy, and surprise you in the first thirty seconds.
- AI is load-bearing — the magic has to come from the models, not from clever scaffolding around them. The multi-agent drama is the product.
- Original — "many worlds, one event-sourced engine, agents that never call each other" is a genuinely different shape from the usual agent-framework playbook.
- Polished — the live theater should be smooth and beautiful, with no rough edges in the demo.
The code and the demo
You can check out the code on GitHub, and the live demo at Hugging Face Spaces.