WorldSeed

Why we included this project

WorldSeed is a runtime for building worlds where many LLM-driven agents interact under explicit rules, private information, and consequences, and where useful behavior emerges from those interactions rather than from a hand-coded workflow. The same engine can host a production room, a simulation, or a fictional setting without a rewrite, so it suits social simulation, game prototypes, and research-style multi-agent pipelines alike. The autoresearch demo shows the payoff: a cohort of specialist agents proposes hypotheses, runs experiments, and peer-reviews each other's papers, and every claim traces back to a commit and a verified result. Because of that traceability, you can watch from above, step into a character, or intervene mid-run and still trust what the agents produced. The world is declared in YAML, agents get asymmetric perception, and the LLM referee is pluggable, so adapting it to your own scenario is straightforward.