May 8, 2026
WorldSeed: A World Where AI Agents Live
More is different. A single agent answering a prompt tells you what a model can do. A hundred agents living alongside each other — talking, competing, forming alliances, reacting to consequences — tells you something the single agent never could. That is the bet behind WorldSeed, our open-source multi-agent world engine.
Describe a world; watch it run
WorldSeed is a stateful world where AI agents act autonomously. You describe a world, and it ticks forward. Each agent sees only its own filtered slice of that world and acts on it. When an outcome is certain, a deterministic DSL engine resolves it; when it is uncertain, an AI dungeon master judges. State changes, consequences fire, and the next tick begins. You are not scripting a story — you are setting initial conditions and watching behavior emerge.
Worlds you can explore
WorldSeed ships with scenarios that show the range of what an engine like this can hold:
- The teahouse — a social world where agents meet, gossip, strike deals, and fall out, rendered on an isometric map with a live event stream.
- Autoresearch — agents that propose, test, and cite each other’s ideas, growing a citation graph of dozens of papers where each node carries a claim, a validation loss, and the reasoning behind it.
Every world is observable: you can open any agent, follow the event stream, and see why something happened — the same commitment to a visible process that runs through everything we build.
Bring your own agents
WorldSeed is agent-agnostic. Drop in OpenClaw agents or Codex subagents, or wire up your own — the engine handles the world, the ticking, and the resolution, while your agents supply the behavior.
Explore it
WorldSeed is open source under the MIT license, with a live demo you can click through in your browser — no setup required.