Multi-Agent LLMs Create Fictional Worlds with Context Compression.

Creating coherent fictional worlds for games or literature is a complex task, and while large language models offer potential for automation, they face challenges like managing vast context, ensuring creative diversity while maintaining consistency, and guaranteeing quality. A new system, AutoWorldBuilder, tackles these issues through a collaborative multi-agent approach. AutoWorldBuilder integrates five key components: a structured concept network for conflict detection, a DAG-based scheduler for grouping tasks by semantic locality, and a four-layer context compression mechanism that reduces token usage by approximately 90%. It also features an iterative review system with specialized "Auditor" agents, which significantly improves the acceptance rate of generated proposals. The system uses a skill-driven agent architecture that allows for zero-code extensions and differentiated temperature configurations. Experiments with various worldbuilding tasks, using GPT-OSS 120B and DeepSeek v3.2, demonstrated a 95% success rate, generating numerous self-consistent concepts per world with zero conflicts in a short timeframe. The architectural patterns, such as layered compression and separation of generation and review, are broadly applicable to other knowledge-intensive, multi-agent LLM applications.