We just open-sourced WorldSeed -- a multi-agent world engine that grew out of MorphMind's backend. AI agents think and act for themselves.
Most multi-agent frameworks focus on getting agents to collaborate on a task. WorldSeed does something different: it lets them live in a world whose rules you define.
You write a YAML config: the spaces, the characters, their goals and secrets, the rules, who can see what. Then you press play. Agents negotiate, deceive, cooperate, betray -- all on their own. No script. No central controller.
What makes it work:
- Deterministic rules engine settles what's computable (movement, items, visibility)
- An LLM Dungeon Master judges what isn't (did they believe the lie? did the mood shift?)
- Real information asymmetry -- each agent only sees what the rules allow
- Everything event-sourced to a replayable log
Two scenarios ship out of the box: a teahouse espionage game and a post-layoff tech company drama. The engine is language-agnostic -- build scenarios in any language.
MIT license.
If you work in social simulation, game design, organizational research, or just want to see what happens when AI agents make their own choices -- take a look.
github.com/AIScientists-Dev/WorldSeed
